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#if !defined(EIGEN_VECTORIZE_AVX) && !defined(EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS)
24#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
27#ifdef EIGEN_VECTORIZE_FMA
28#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
29#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
33#if ((defined EIGEN_VECTORIZE_AVX) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_MINGW) && (__GXX_ABI_VERSION < 1004)) || EIGEN_OS_QNX
39typedef eigen_packet_wrapper<__m128> Packet4f;
40typedef eigen_packet_wrapper<__m128d> Packet2d;
42typedef __m128 Packet4f;
43typedef __m128d Packet2d;
46typedef eigen_packet_wrapper<__m128i, 0> Packet4i;
47typedef eigen_packet_wrapper<__m128i, 1> Packet16b;
49template<>
struct is_arithmetic<__m128> {
enum { value =
true }; };
50template<>
struct is_arithmetic<__m128i> {
enum { value =
true }; };
51template<>
struct is_arithmetic<__m128d> {
enum { value =
true }; };
52template<>
struct is_arithmetic<Packet4i> {
enum { value =
true }; };
53template<>
struct is_arithmetic<Packet16b> {
enum { value =
true }; };
55template<
int p,
int q,
int r,
int s>
57 enum { mask = (s)<<6|(r)<<4|(q)<<2|(p) };
61#define vec4f_swizzle1(v,p,q,r,s) \
62 Packet4f(_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), (shuffle_mask<p,q,r,s>::mask))))
64#define vec4i_swizzle1(v,p,q,r,s) \
65 Packet4i(_mm_shuffle_epi32( v, (shuffle_mask<p,q,r,s>::mask)))
67#define vec2d_swizzle1(v,p,q) \
68 Packet2d(_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), (shuffle_mask<2*p,2*p+1,2*q,2*q+1>::mask))))
70#define vec4f_swizzle2(a,b,p,q,r,s) \
71 Packet4f(_mm_shuffle_ps( (a), (b), (shuffle_mask<p,q,r,s>::mask)))
73#define vec4i_swizzle2(a,b,p,q,r,s) \
74 Packet4i(_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), (shuffle_mask<p,q,r,s>::mask)))))
76EIGEN_STRONG_INLINE Packet4f vec4f_movelh(
const Packet4f& a,
const Packet4f& b)
78 return Packet4f(_mm_movelh_ps(a,b));
80EIGEN_STRONG_INLINE Packet4f vec4f_movehl(
const Packet4f& a,
const Packet4f& b)
82 return Packet4f(_mm_movehl_ps(a,b));
84EIGEN_STRONG_INLINE Packet4f vec4f_unpacklo(
const Packet4f& a,
const Packet4f& b)
86 return Packet4f(_mm_unpacklo_ps(a,b));
88EIGEN_STRONG_INLINE Packet4f vec4f_unpackhi(
const Packet4f& a,
const Packet4f& b)
90 return Packet4f(_mm_unpackhi_ps(a,b));
92#define vec4f_duplane(a,p) \
93 vec4f_swizzle2(a,a,p,p,p,p)
95#define vec2d_swizzle2(a,b,mask) \
96 Packet2d(_mm_shuffle_pd(a,b,mask))
98EIGEN_STRONG_INLINE Packet2d vec2d_unpacklo(
const Packet2d& a,
const Packet2d& b)
100 return Packet2d(_mm_unpacklo_pd(a,b));
102EIGEN_STRONG_INLINE Packet2d vec2d_unpackhi(
const Packet2d& a,
const Packet2d& b)
104 return Packet2d(_mm_unpackhi_pd(a,b));
106#define vec2d_duplane(a,p) \
107 vec2d_swizzle2(a,a,(p<<1)|p)
109#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
110 const Packet4f p4f_##NAME = pset1<Packet4f>(X)
112#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
113 const Packet2d p2d_##NAME = pset1<Packet2d>(X)
115#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
116 const Packet4f p4f_##NAME = pset1frombits<Packet4f>(X)
118#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
119 const Packet4i p4i_##NAME = pset1<Packet4i>(X)
124#ifndef EIGEN_VECTORIZE_AVX
126struct packet_traits<float> : default_packet_traits {
127 typedef Packet4f type;
128 typedef Packet4f half;
137 HasSin = EIGEN_FAST_MATH,
138 HasCos = EIGEN_FAST_MATH,
147 HasTanh = EIGEN_FAST_MATH,
148 HasErf = EIGEN_FAST_MATH,
152#ifdef EIGEN_VECTORIZE_SSE4_1
159struct packet_traits<double> : default_packet_traits {
160 typedef Packet2d type;
161 typedef Packet2d half;
177#ifdef EIGEN_VECTORIZE_SSE4_1
184template<>
struct packet_traits<int> : default_packet_traits
186 typedef Packet4i type;
187 typedef Packet4i half;
198template<>
struct packet_traits<bool> : default_packet_traits
200 typedef Packet16b type;
201 typedef Packet16b half;
222template<>
struct unpacket_traits<Packet4f> {
224 typedef Packet4f half;
225 typedef Packet4i integer_packet;
226 enum {size=4, alignment=
Aligned16, vectorizable=
true, masked_load_available=
false, masked_store_available=
false};
228template<>
struct unpacket_traits<Packet2d> {
230 typedef Packet2d half;
231 enum {size=2, alignment=
Aligned16, vectorizable=
true, masked_load_available=
false, masked_store_available=
false};
233template<>
struct unpacket_traits<Packet4i> {
235 typedef Packet4i half;
236 enum {size=4, alignment=
Aligned16, vectorizable=
false, masked_load_available=
false, masked_store_available=
false};
238template<>
struct unpacket_traits<Packet16b> {
240 typedef Packet16b half;
241 enum {size=16, alignment=
Aligned16, vectorizable=
true, masked_load_available=
false, masked_store_available=
false};
244#ifndef EIGEN_VECTORIZE_AVX
245template<>
struct scalar_div_cost<float,true> {
enum { value = 7 }; };
246template<>
struct scalar_div_cost<double,true> {
enum { value = 8 }; };
249#if EIGEN_COMP_MSVC==1500
253template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(
const float& from) {
return _mm_set_ps(from,from,from,from); }
254template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(
const double& from) {
return _mm_set_pd(from,from); }
255template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(
const int& from) {
return _mm_set_epi32(from,from,from,from); }
257template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(
const float& from) {
return _mm_set_ps1(from); }
258template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(
const double& from) {
return _mm_set1_pd(from); }
259template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(
const int& from) {
return _mm_set1_epi32(from); }
261template<> EIGEN_STRONG_INLINE Packet16b pset1<Packet16b>(
const bool& from) {
return _mm_set1_epi8(
static_cast<char>(from)); }
263template<> EIGEN_STRONG_INLINE Packet4f pset1frombits<Packet4f>(
unsigned int from) {
return _mm_castsi128_ps(pset1<Packet4i>(from)); }
264template<> EIGEN_STRONG_INLINE Packet2d pset1frombits<Packet2d>(uint64_t from) {
return _mm_castsi128_pd(_mm_set1_epi64x(from)); }
266template<> EIGEN_STRONG_INLINE Packet4f peven_mask(
const Packet4f& ) {
return _mm_castsi128_ps(_mm_set_epi32(0, -1, 0, -1)); }
267template<> EIGEN_STRONG_INLINE Packet4i peven_mask(
const Packet4i& ) {
return _mm_set_epi32(0, -1, 0, -1); }
268template<> EIGEN_STRONG_INLINE Packet2d peven_mask(
const Packet2d& ) {
return _mm_castsi128_pd(_mm_set_epi32(0, 0, -1, -1)); }
270template<> EIGEN_STRONG_INLINE Packet4f pzero(
const Packet4f& ) {
return _mm_setzero_ps(); }
271template<> EIGEN_STRONG_INLINE Packet2d pzero(
const Packet2d& ) {
return _mm_setzero_pd(); }
272template<> EIGEN_STRONG_INLINE Packet4i pzero(
const Packet4i& ) {
return _mm_setzero_si128(); }
279#if EIGEN_COMP_GNUC_STRICT && (!defined __AVX__)
280template<> EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(
const float *from) {
281 return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0);
285template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(
const float& a) {
return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
286template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(
const double& a) {
return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
287template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(
const int& a) {
return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); }
289template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_add_ps(a,b); }
290template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_add_pd(a,b); }
291template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_add_epi32(a,b); }
293template<> EIGEN_STRONG_INLINE Packet16b padd<Packet16b>(
const Packet16b& a,
const Packet16b& b) {
return _mm_or_si128(a,b); }
295template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_sub_ps(a,b); }
296template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_sub_pd(a,b); }
297template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_sub_epi32(a,b); }
298template<> EIGEN_STRONG_INLINE Packet16b psub<Packet16b>(
const Packet16b& a,
const Packet16b& b) {
return _mm_xor_si128(a,b); }
300template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(
const Packet4f& a,
const Packet4f& b);
301template<> EIGEN_STRONG_INLINE Packet4f paddsub<Packet4f>(
const Packet4f& a,
const Packet4f& b)
303#ifdef EIGEN_VECTORIZE_SSE3
304 return _mm_addsub_ps(a,b);
306 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x0,0x80000000,0x0));
307 return padd(a, pxor(mask, b));
311template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(
const Packet2d& ,
const Packet2d& );
312template<> EIGEN_STRONG_INLINE Packet2d paddsub<Packet2d>(
const Packet2d& a,
const Packet2d& b)
314#ifdef EIGEN_VECTORIZE_SSE3
315 return _mm_addsub_pd(a,b);
317 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x0));
318 return padd(a, pxor(mask, b));
322template<> EIGEN_STRONG_INLINE Packet4f pnegate(
const Packet4f& a)
324 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
325 return _mm_xor_ps(a,mask);
327template<> EIGEN_STRONG_INLINE Packet2d pnegate(
const Packet2d& a)
329 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000));
330 return _mm_xor_pd(a,mask);
332template<> EIGEN_STRONG_INLINE Packet4i pnegate(
const Packet4i& a)
334 return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a);
337template<> EIGEN_STRONG_INLINE Packet16b pnegate(
const Packet16b& a)
339 return psub(pset1<Packet16b>(
false), a);
342template<> EIGEN_STRONG_INLINE Packet4f pconj(
const Packet4f& a) {
return a; }
343template<> EIGEN_STRONG_INLINE Packet2d pconj(
const Packet2d& a) {
return a; }
344template<> EIGEN_STRONG_INLINE Packet4i pconj(
const Packet4i& a) {
return a; }
346template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_mul_ps(a,b); }
347template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_mul_pd(a,b); }
348template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(
const Packet4i& a,
const Packet4i& b)
350#ifdef EIGEN_VECTORIZE_SSE4_1
351 return _mm_mullo_epi32(a,b);
354 return vec4i_swizzle1(
357 _mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2),
358 vec4i_swizzle1(b,1,0,3,2)),
364template<> EIGEN_STRONG_INLINE Packet16b pmul<Packet16b>(
const Packet16b& a,
const Packet16b& b) {
return _mm_and_si128(a,b); }
366template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_div_ps(a,b); }
367template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_div_pd(a,b); }
370template<> EIGEN_STRONG_INLINE Packet4i pmadd(
const Packet4i& a,
const Packet4i& b,
const Packet4i& c) {
return padd(pmul(a,b), c); }
371#ifdef EIGEN_VECTORIZE_FMA
372template<> EIGEN_STRONG_INLINE Packet4f pmadd(
const Packet4f& a,
const Packet4f& b,
const Packet4f& c) {
return _mm_fmadd_ps(a,b,c); }
373template<> EIGEN_STRONG_INLINE Packet2d pmadd(
const Packet2d& a,
const Packet2d& b,
const Packet2d& c) {
return _mm_fmadd_pd(a,b,c); }
376#ifdef EIGEN_VECTORIZE_SSE4_1
377template<> EIGEN_DEVICE_FUNC
inline Packet4f pselect(
const Packet4f& mask,
const Packet4f& a,
const Packet4f& b) {
378 return _mm_blendv_ps(b,a,mask);
381template<> EIGEN_DEVICE_FUNC
inline Packet4i pselect(
const Packet4i& mask,
const Packet4i& a,
const Packet4i& b) {
382 return _mm_castps_si128(_mm_blendv_ps(_mm_castsi128_ps(b),_mm_castsi128_ps(a),_mm_castsi128_ps(mask)));
385template<> EIGEN_DEVICE_FUNC
inline Packet2d pselect(
const Packet2d& mask,
const Packet2d& a,
const Packet2d& b) {
return _mm_blendv_pd(b,a,mask); }
387template<> EIGEN_DEVICE_FUNC
inline Packet16b pselect(
const Packet16b& mask,
const Packet16b& a,
const Packet16b& b) {
388 return _mm_blendv_epi8(b,a,mask);
391template<> EIGEN_DEVICE_FUNC
inline Packet16b pselect(
const Packet16b& mask,
const Packet16b& a,
const Packet16b& b) {
392 Packet16b a_part = _mm_and_si128(mask, a);
393 Packet16b b_part = _mm_andnot_si128(mask, b);
394 return _mm_or_si128(a_part, b_part);
398template<> EIGEN_STRONG_INLINE Packet4i ptrue<Packet4i>(
const Packet4i& a) {
return _mm_cmpeq_epi32(a, a); }
399template<> EIGEN_STRONG_INLINE Packet16b ptrue<Packet16b>(
const Packet16b& a) {
return _mm_cmpeq_epi8(a, a); }
400template<> EIGEN_STRONG_INLINE Packet4f
401ptrue<Packet4f>(
const Packet4f& a) {
402 Packet4i b = _mm_castps_si128(a);
403 return _mm_castsi128_ps(_mm_cmpeq_epi32(b, b));
405template<> EIGEN_STRONG_INLINE Packet2d
406ptrue<Packet2d>(
const Packet2d& a) {
407 Packet4i b = _mm_castpd_si128(a);
408 return _mm_castsi128_pd(_mm_cmpeq_epi32(b, b));
412template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_and_ps(a,b); }
413template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_and_pd(a,b); }
414template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_and_si128(a,b); }
415template<> EIGEN_STRONG_INLINE Packet16b pand<Packet16b>(
const Packet16b& a,
const Packet16b& b) {
return _mm_and_si128(a,b); }
417template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_or_ps(a,b); }
418template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_or_pd(a,b); }
419template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_or_si128(a,b); }
420template<> EIGEN_STRONG_INLINE Packet16b por<Packet16b>(
const Packet16b& a,
const Packet16b& b) {
return _mm_or_si128(a,b); }
422template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_xor_ps(a,b); }
423template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_xor_pd(a,b); }
424template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_xor_si128(a,b); }
425template<> EIGEN_STRONG_INLINE Packet16b pxor<Packet16b>(
const Packet16b& a,
const Packet16b& b) {
return _mm_xor_si128(a,b); }
427template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_andnot_ps(b,a); }
428template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_andnot_pd(b,a); }
429template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_andnot_si128(b,a); }
431template<> EIGEN_STRONG_INLINE Packet4f pcmp_le(
const Packet4f& a,
const Packet4f& b) {
return _mm_cmple_ps(a,b); }
432template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(
const Packet4f& a,
const Packet4f& b) {
return _mm_cmplt_ps(a,b); }
433template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(
const Packet4f& a,
const Packet4f& b) {
return _mm_cmpnge_ps(a,b); }
434template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(
const Packet4f& a,
const Packet4f& b) {
return _mm_cmpeq_ps(a,b); }
436template<> EIGEN_STRONG_INLINE Packet2d pcmp_le(
const Packet2d& a,
const Packet2d& b) {
return _mm_cmple_pd(a,b); }
437template<> EIGEN_STRONG_INLINE Packet2d pcmp_lt(
const Packet2d& a,
const Packet2d& b) {
return _mm_cmplt_pd(a,b); }
438template<> EIGEN_STRONG_INLINE Packet2d pcmp_lt_or_nan(
const Packet2d& a,
const Packet2d& b) {
return _mm_cmpnge_pd(a,b); }
439template<> EIGEN_STRONG_INLINE Packet2d pcmp_eq(
const Packet2d& a,
const Packet2d& b) {
return _mm_cmpeq_pd(a,b); }
441template<> EIGEN_STRONG_INLINE Packet4i pcmp_lt(
const Packet4i& a,
const Packet4i& b) {
return _mm_cmplt_epi32(a,b); }
442template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(
const Packet4i& a,
const Packet4i& b) {
return _mm_cmpeq_epi32(a,b); }
443template<> EIGEN_STRONG_INLINE Packet16b pcmp_eq(
const Packet16b& a,
const Packet16b& b) {
return _mm_cmpeq_epi8(a,b); }
444template<> EIGEN_STRONG_INLINE Packet4i pcmp_le(
const Packet4i& a,
const Packet4i& b) {
return por(pcmp_lt(a,b), pcmp_eq(a,b)); }
446template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
447#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
452 #ifdef EIGEN_VECTORIZE_AVX
454 asm(
"vminps %[a], %[b], %[res]" : [res]
"=x" (res) : [a]
"x" (a), [b]
"x" (b));
457 asm(
"minps %[a], %[res]" : [res]
"+x" (res) : [a]
"x" (a));
462 return _mm_min_ps(b, a);
465template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
466#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
471 #ifdef EIGEN_VECTORIZE_AVX
473 asm(
"vminpd %[a], %[b], %[res]" : [res]
"=x" (res) : [a]
"x" (a), [b]
"x" (b));
476 asm(
"minpd %[a], %[res]" : [res]
"+x" (res) : [a]
"x" (a));
481 return _mm_min_pd(b, a);
484template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(
const Packet4i& a,
const Packet4i& b)
486#ifdef EIGEN_VECTORIZE_SSE4_1
487 return _mm_min_epi32(a,b);
490 Packet4i mask = _mm_cmplt_epi32(a,b);
491 return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
496template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
497#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
502 #ifdef EIGEN_VECTORIZE_AVX
504 asm(
"vmaxps %[a], %[b], %[res]" : [res]
"=x" (res) : [a]
"x" (a), [b]
"x" (b));
507 asm(
"maxps %[a], %[res]" : [res]
"+x" (res) : [a]
"x" (a));
512 return _mm_max_ps(b, a);
515template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
516#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
521 #ifdef EIGEN_VECTORIZE_AVX
523 asm(
"vmaxpd %[a], %[b], %[res]" : [res]
"=x" (res) : [a]
"x" (a), [b]
"x" (b));
526 asm(
"maxpd %[a], %[res]" : [res]
"+x" (res) : [a]
"x" (a));
531 return _mm_max_pd(b, a);
534template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(
const Packet4i& a,
const Packet4i& b)
536#ifdef EIGEN_VECTORIZE_SSE4_1
537 return _mm_max_epi32(a,b);
540 Packet4i mask = _mm_cmpgt_epi32(a,b);
541 return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
545template <
typename Packet,
typename Op>
546EIGEN_STRONG_INLINE Packet pminmax_propagate_numbers(
const Packet& a,
const Packet& b, Op op) {
549 Packet not_nan_mask_a = pcmp_eq(a, a);
551 return pselect<Packet>(not_nan_mask_a, m, b);
554template <
typename Packet,
typename Op>
555EIGEN_STRONG_INLINE Packet pminmax_propagate_nan(
const Packet& a,
const Packet& b, Op op) {
558 Packet not_nan_mask_a = pcmp_eq(a, a);
560 return pselect<Packet>(not_nan_mask_a, m, a);
565EIGEN_STRONG_INLINE Packet4f pmin<PropagateNumbers, Packet4f>(
const Packet4f& a,
const Packet4f& b) {
566 return pminmax_propagate_numbers(a, b, pmin<Packet4f>);
569EIGEN_STRONG_INLINE Packet2d pmin<PropagateNumbers, Packet2d>(
const Packet2d& a,
const Packet2d& b) {
570 return pminmax_propagate_numbers(a, b, pmin<Packet2d>);
573EIGEN_STRONG_INLINE Packet4f pmax<PropagateNumbers, Packet4f>(
const Packet4f& a,
const Packet4f& b) {
574 return pminmax_propagate_numbers(a, b, pmax<Packet4f>);
577EIGEN_STRONG_INLINE Packet2d pmax<PropagateNumbers, Packet2d>(
const Packet2d& a,
const Packet2d& b) {
578 return pminmax_propagate_numbers(a, b, pmax<Packet2d>);
581EIGEN_STRONG_INLINE Packet4f pmin<PropagateNaN, Packet4f>(
const Packet4f& a,
const Packet4f& b) {
582 return pminmax_propagate_nan(a, b, pmin<Packet4f>);
585EIGEN_STRONG_INLINE Packet2d pmin<PropagateNaN, Packet2d>(
const Packet2d& a,
const Packet2d& b) {
586 return pminmax_propagate_nan(a, b, pmin<Packet2d>);
589EIGEN_STRONG_INLINE Packet4f pmax<PropagateNaN, Packet4f>(
const Packet4f& a,
const Packet4f& b) {
590 return pminmax_propagate_nan(a, b, pmax<Packet4f>);
593EIGEN_STRONG_INLINE Packet2d pmax<PropagateNaN, Packet2d>(
const Packet2d& a,
const Packet2d& b) {
594 return pminmax_propagate_nan(a, b, pmax<Packet2d>);
597template<
int N> EIGEN_STRONG_INLINE Packet4i parithmetic_shift_right(
const Packet4i& a) {
return _mm_srai_epi32(a,N); }
598template<
int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_right (
const Packet4i& a) {
return _mm_srli_epi32(a,N); }
599template<
int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_left (
const Packet4i& a) {
return _mm_slli_epi32(a,N); }
601template<> EIGEN_STRONG_INLINE Packet4f pabs(
const Packet4f& a)
603 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
604 return _mm_and_ps(a,mask);
606template<> EIGEN_STRONG_INLINE Packet2d pabs(
const Packet2d& a)
608 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF));
609 return _mm_and_pd(a,mask);
611template<> EIGEN_STRONG_INLINE Packet4i pabs(
const Packet4i& a)
613 #ifdef EIGEN_VECTORIZE_SSSE3
614 return _mm_abs_epi32(a);
616 Packet4i aux = _mm_srai_epi32(a,31);
617 return _mm_sub_epi32(_mm_xor_si128(a,aux),aux);
621#ifdef EIGEN_VECTORIZE_SSE4_1
622template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(
const Packet4f& a)
625 const Packet4f mask = pset1frombits<Packet4f>(0x80000000u);
626 const Packet4f prev0dot5 = pset1frombits<Packet4f>(0x3EFFFFFFu);
627 return _mm_round_ps(padd(por(pand(a, mask), prev0dot5), a), _MM_FROUND_TO_ZERO);
630template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(
const Packet2d& a)
632 const Packet2d mask = _mm_castsi128_pd(_mm_set_epi64x(0x8000000000000000ull, 0x8000000000000000ull));
633 const Packet2d prev0dot5 = _mm_castsi128_pd(_mm_set_epi64x(0x3FDFFFFFFFFFFFFFull, 0x3FDFFFFFFFFFFFFFull));
634 return _mm_round_pd(padd(por(pand(a, mask), prev0dot5), a), _MM_FROUND_TO_ZERO);
637template<> EIGEN_STRONG_INLINE Packet4f print<Packet4f>(
const Packet4f& a) {
return _mm_round_ps(a, _MM_FROUND_CUR_DIRECTION); }
638template<> EIGEN_STRONG_INLINE Packet2d print<Packet2d>(
const Packet2d& a) {
return _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); }
640template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(
const Packet4f& a) {
return _mm_ceil_ps(a); }
641template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(
const Packet2d& a) {
return _mm_ceil_pd(a); }
643template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(
const Packet4f& a) {
return _mm_floor_ps(a); }
644template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(
const Packet2d& a) {
return _mm_floor_pd(a); }
646template<> EIGEN_STRONG_INLINE Packet4f print(
const Packet4f& a) {
648 const Packet4f limit = pset1<Packet4f>(
static_cast<float>(1<<23));
649 const Packet4f abs_a = pabs(a);
650 Packet4f r = padd(abs_a, limit);
652 EIGEN_OPTIMIZATION_BARRIER(r);
655 r = pselect(pcmp_lt(abs_a, limit),
656 pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
660template<> EIGEN_STRONG_INLINE Packet2d print(
const Packet2d& a) {
662 const Packet2d limit = pset1<Packet2d>(
static_cast<double>(1ull<<52));
663 const Packet2d abs_a = pabs(a);
664 Packet2d r = padd(abs_a, limit);
666 EIGEN_OPTIMIZATION_BARRIER(r);
669 r = pselect(pcmp_lt(abs_a, limit),
670 pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
674template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(
const Packet4f& a)
676 const Packet4f cst_1 = pset1<Packet4f>(1.0f);
677 Packet4f tmp = print<Packet4f>(a);
679 Packet4f mask = _mm_cmpgt_ps(tmp, a);
680 mask = pand(mask, cst_1);
681 return psub(tmp, mask);
684template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(
const Packet2d& a)
686 const Packet2d cst_1 = pset1<Packet2d>(1.0);
687 Packet2d tmp = print<Packet2d>(a);
689 Packet2d mask = _mm_cmpgt_pd(tmp, a);
690 mask = pand(mask, cst_1);
691 return psub(tmp, mask);
694template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(
const Packet4f& a)
696 const Packet4f cst_1 = pset1<Packet4f>(1.0f);
697 Packet4f tmp = print<Packet4f>(a);
699 Packet4f mask = _mm_cmplt_ps(tmp, a);
700 mask = pand(mask, cst_1);
701 return padd(tmp, mask);
704template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(
const Packet2d& a)
706 const Packet2d cst_1 = pset1<Packet2d>(1.0);
707 Packet2d tmp = print<Packet2d>(a);
709 Packet2d mask = _mm_cmplt_pd(tmp, a);
710 mask = pand(mask, cst_1);
711 return padd(tmp, mask);
715template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(
const float* from) { EIGEN_DEBUG_ALIGNED_LOAD
return _mm_load_ps(from); }
716template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(
const double* from) { EIGEN_DEBUG_ALIGNED_LOAD
return _mm_load_pd(from); }
717template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(
const int* from) { EIGEN_DEBUG_ALIGNED_LOAD
return _mm_load_si128(
reinterpret_cast<const __m128i*
>(from)); }
718template<> EIGEN_STRONG_INLINE Packet16b pload<Packet16b>(
const bool* from) { EIGEN_DEBUG_ALIGNED_LOAD
return _mm_load_si128(
reinterpret_cast<const __m128i*
>(from)); }
721 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(
const float* from) {
722 EIGEN_DEBUG_UNALIGNED_LOAD
723 #if (EIGEN_COMP_MSVC==1600)
726 __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (
const __m64*)(from));
727 res = _mm_loadh_pi(res, (
const __m64*)(from+2));
730 return _mm_loadu_ps(from);
736template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(
const float* from)
738 EIGEN_DEBUG_UNALIGNED_LOAD
739 return _mm_loadu_ps(from);
743template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(
const double* from)
745 EIGEN_DEBUG_UNALIGNED_LOAD
746 return _mm_loadu_pd(from);
748template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(
const int* from)
750 EIGEN_DEBUG_UNALIGNED_LOAD
751 return _mm_loadu_si128(
reinterpret_cast<const __m128i*
>(from));
753template<> EIGEN_STRONG_INLINE Packet16b ploadu<Packet16b>(
const bool* from) {
754 EIGEN_DEBUG_UNALIGNED_LOAD
755 return _mm_loadu_si128(
reinterpret_cast<const __m128i*
>(from));
759template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(
const float* from)
761 return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(
reinterpret_cast<const double*
>(from))), 0, 0, 1, 1);
763template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(
const double* from)
764{
return pset1<Packet2d>(from[0]); }
765template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(
const int* from)
768 tmp = _mm_loadl_epi64(
reinterpret_cast<const __m128i*
>(from));
769 return vec4i_swizzle1(tmp, 0, 0, 1, 1);
774template<> EIGEN_STRONG_INLINE Packet16b ploaddup<Packet16b>(
const bool* from)
776 __m128i tmp = _mm_castpd_si128(pload1<Packet2d>(
reinterpret_cast<const double*
>(from)));
777 return _mm_unpacklo_epi8(tmp, tmp);
782template<> EIGEN_STRONG_INLINE Packet16b
783ploadquad<Packet16b>(
const bool* from) {
784 __m128i tmp = _mm_castps_si128(pload1<Packet4f>(
reinterpret_cast<const float*
>(from)));
785 tmp = _mm_unpacklo_epi8(tmp, tmp);
786 return _mm_unpacklo_epi16(tmp, tmp);
789template<> EIGEN_STRONG_INLINE
void pstore<float>(
float* to,
const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); }
790template<> EIGEN_STRONG_INLINE
void pstore<double>(
double* to,
const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); }
791template<> EIGEN_STRONG_INLINE
void pstore<int>(
int* to,
const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(
reinterpret_cast<__m128i*
>(to), from); }
792template<> EIGEN_STRONG_INLINE
void pstore<bool>(
bool* to,
const Packet16b& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(
reinterpret_cast<__m128i*
>(to), from); }
794template<> EIGEN_STRONG_INLINE
void pstoreu<double>(
double* to,
const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); }
795template<> EIGEN_STRONG_INLINE
void pstoreu<float>(
float* to,
const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); }
796template<> EIGEN_STRONG_INLINE
void pstoreu<int>(
int* to,
const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(
reinterpret_cast<__m128i*
>(to), from); }
797template<> EIGEN_STRONG_INLINE
void pstoreu<bool>(
bool* to,
const Packet16b& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_storeu_si128(
reinterpret_cast<__m128i*
>(to), from); }
799template<> EIGEN_DEVICE_FUNC
inline Packet4f pgather<float, Packet4f>(
const float* from,
Index stride)
801 return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
803template<> EIGEN_DEVICE_FUNC
inline Packet2d pgather<double, Packet2d>(
const double* from,
Index stride)
805 return _mm_set_pd(from[1*stride], from[0*stride]);
807template<> EIGEN_DEVICE_FUNC
inline Packet4i pgather<int, Packet4i>(
const int* from,
Index stride)
809 return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
812template<> EIGEN_DEVICE_FUNC
inline Packet16b pgather<bool, Packet16b>(
const bool* from,
Index stride)
814 return _mm_set_epi8(from[15*stride], from[14*stride], from[13*stride], from[12*stride],
815 from[11*stride], from[10*stride], from[9*stride], from[8*stride],
816 from[7*stride], from[6*stride], from[5*stride], from[4*stride],
817 from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
820template<> EIGEN_DEVICE_FUNC
inline void pscatter<float, Packet4f>(
float* to,
const Packet4f& from,
Index stride)
822 to[stride*0] = _mm_cvtss_f32(from);
823 to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1));
824 to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2));
825 to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3));
827template<> EIGEN_DEVICE_FUNC
inline void pscatter<double, Packet2d>(
double* to,
const Packet2d& from,
Index stride)
829 to[stride*0] = _mm_cvtsd_f64(from);
830 to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1));
832template<> EIGEN_DEVICE_FUNC
inline void pscatter<int, Packet4i>(
int* to,
const Packet4i& from,
Index stride)
834 to[stride*0] = _mm_cvtsi128_si32(from);
835 to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1));
836 to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2));
837 to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3));
839template<> EIGEN_DEVICE_FUNC
inline void pscatter<bool, Packet16b>(
bool* to,
const Packet16b& from,
Index stride)
841 to[4*stride*0] = _mm_cvtsi128_si32(from);
842 to[4*stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1));
843 to[4*stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2));
844 to[4*stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3));
849template<> EIGEN_STRONG_INLINE
void pstore1<Packet4f>(
float* to,
const float& a)
851 Packet4f pa = _mm_set_ss(a);
852 pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0)));
855template<> EIGEN_STRONG_INLINE
void pstore1<Packet2d>(
double* to,
const double& a)
857 Packet2d pa = _mm_set_sd(a);
858 pstore(to, Packet2d(vec2d_swizzle1(pa,0,0)));
861#if EIGEN_COMP_PGI && EIGEN_COMP_PGI < 1900
862typedef const void * SsePrefetchPtrType;
864typedef const char * SsePrefetchPtrType;
867#ifndef EIGEN_VECTORIZE_AVX
868template<> EIGEN_STRONG_INLINE
void prefetch<float>(
const float* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
869template<> EIGEN_STRONG_INLINE
void prefetch<double>(
const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
870template<> EIGEN_STRONG_INLINE
void prefetch<int>(
const int* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
873#if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64
876template<> EIGEN_STRONG_INLINE
float pfirst<Packet4f>(
const Packet4f& a) {
return a.m128_f32[0]; }
877template<> EIGEN_STRONG_INLINE
double pfirst<Packet2d>(
const Packet2d& a) {
return a.m128d_f64[0]; }
878template<> EIGEN_STRONG_INLINE
int pfirst<Packet4i>(
const Packet4i& a) {
int x = _mm_cvtsi128_si32(a);
return x; }
879#elif EIGEN_COMP_MSVC_STRICT
881template<> EIGEN_STRONG_INLINE
float pfirst<Packet4f>(
const Packet4f& a) {
float x = _mm_cvtss_f32(a);
return x; }
882template<> EIGEN_STRONG_INLINE
double pfirst<Packet2d>(
const Packet2d& a) {
double x = _mm_cvtsd_f64(a);
return x; }
883template<> EIGEN_STRONG_INLINE
int pfirst<Packet4i>(
const Packet4i& a) {
int x = _mm_cvtsi128_si32(a);
return x; }
885template<> EIGEN_STRONG_INLINE
float pfirst<Packet4f>(
const Packet4f& a) {
return _mm_cvtss_f32(a); }
886template<> EIGEN_STRONG_INLINE
double pfirst<Packet2d>(
const Packet2d& a) {
return _mm_cvtsd_f64(a); }
887template<> EIGEN_STRONG_INLINE
int pfirst<Packet4i>(
const Packet4i& a) {
return _mm_cvtsi128_si32(a); }
889template<> EIGEN_STRONG_INLINE
bool pfirst<Packet16b>(
const Packet16b& a) {
int x = _mm_cvtsi128_si32(a);
return static_cast<bool>(x & 1); }
891template<> EIGEN_STRONG_INLINE Packet4f preverse(
const Packet4f& a) {
return _mm_shuffle_ps(a,a,0x1B); }
892template<> EIGEN_STRONG_INLINE Packet2d preverse(
const Packet2d& a) {
return _mm_shuffle_pd(a,a,0x1); }
893template<> EIGEN_STRONG_INLINE Packet4i preverse(
const Packet4i& a) {
return _mm_shuffle_epi32(a,0x1B); }
894template<> EIGEN_STRONG_INLINE Packet16b preverse(
const Packet16b& a) {
895#ifdef EIGEN_VECTORIZE_SSSE3
896 __m128i mask = _mm_set_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
897 return _mm_shuffle_epi8(a, mask);
899 Packet16b tmp = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 1, 2, 3));
900 tmp = _mm_shufflehi_epi16(_mm_shufflelo_epi16(tmp, _MM_SHUFFLE(2, 3, 0, 1)), _MM_SHUFFLE(2, 3, 0, 1));
901 return _mm_or_si128(_mm_slli_epi16(tmp, 8), _mm_srli_epi16(tmp, 8));
905template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(
const Packet4f& a, Packet4f& exponent) {
906 return pfrexp_generic(a,exponent);
912Packet2d pfrexp_generic_get_biased_exponent(
const Packet2d& a) {
913 const Packet2d cst_exp_mask = pset1frombits<Packet2d>(
static_cast<uint64_t
>(0x7ff0000000000000ull));
914 __m128i a_expo = _mm_srli_epi64(_mm_castpd_si128(pand(a, cst_exp_mask)), 52);
915 return _mm_cvtepi32_pd(vec4i_swizzle1(a_expo, 0, 2, 1, 3));
918template<> EIGEN_STRONG_INLINE Packet2d pfrexp<Packet2d>(
const Packet2d& a, Packet2d& exponent) {
919 return pfrexp_generic(a, exponent);
922template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(
const Packet4f& a,
const Packet4f& exponent) {
923 return pldexp_generic(a,exponent);
928template<> EIGEN_STRONG_INLINE Packet2d pldexp<Packet2d>(
const Packet2d& a,
const Packet2d& exponent) {
930 const Packet2d max_exponent = pset1<Packet2d>(2099.0);
931 const Packet2d e = pmin(pmax(exponent, pnegate(max_exponent)), max_exponent);
934 const Packet4i ei = vec4i_swizzle1(_mm_cvtpd_epi32(e), 0, 3, 1, 3);
937 const Packet4i bias = _mm_set_epi32(0, 1023, 0, 1023);
938 Packet4i b = parithmetic_shift_right<2>(ei);
939 Packet2d c = _mm_castsi128_pd(_mm_slli_epi64(padd(b, bias), 52));
940 Packet2d out = pmul(pmul(pmul(a, c), c), c);
941 b = psub(psub(psub(ei, b), b), b);
942 c = _mm_castsi128_pd(_mm_slli_epi64(padd(b, bias), 52));
949template<> EIGEN_STRONG_INLINE
void
950pbroadcast4<Packet4f>(
const float *a,
951 Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
953 a3 = pload<Packet4f>(a);
954 a0 = vec4f_swizzle1(a3, 0,0,0,0);
955 a1 = vec4f_swizzle1(a3, 1,1,1,1);
956 a2 = vec4f_swizzle1(a3, 2,2,2,2);
957 a3 = vec4f_swizzle1(a3, 3,3,3,3);
959template<> EIGEN_STRONG_INLINE
void
960pbroadcast4<Packet2d>(
const double *a,
961 Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
963#ifdef EIGEN_VECTORIZE_SSE3
964 a0 = _mm_loaddup_pd(a+0);
965 a1 = _mm_loaddup_pd(a+1);
966 a2 = _mm_loaddup_pd(a+2);
967 a3 = _mm_loaddup_pd(a+3);
969 a1 = pload<Packet2d>(a);
970 a0 = vec2d_swizzle1(a1, 0,0);
971 a1 = vec2d_swizzle1(a1, 1,1);
972 a3 = pload<Packet2d>(a+2);
973 a2 = vec2d_swizzle1(a3, 0,0);
974 a3 = vec2d_swizzle1(a3, 1,1);
979EIGEN_STRONG_INLINE
void punpackp(Packet4f* vecs)
981 vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55));
982 vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA));
983 vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF));
984 vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00));
987template<> EIGEN_STRONG_INLINE
float predux<Packet4f>(
const Packet4f& a)
995 Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
996 return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
1000template<> EIGEN_STRONG_INLINE
double predux<Packet2d>(
const Packet2d& a)
1007 return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
1011#ifdef EIGEN_VECTORIZE_SSSE3
1012template<> EIGEN_STRONG_INLINE
int predux<Packet4i>(
const Packet4i& a)
1014 Packet4i tmp0 = _mm_hadd_epi32(a,a);
1015 return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0));
1019template<> EIGEN_STRONG_INLINE
int predux<Packet4i>(
const Packet4i& a)
1021 Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
1022 return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1));
1026template<> EIGEN_STRONG_INLINE
bool predux<Packet16b>(
const Packet16b& a) {
1027 Packet4i tmp = _mm_or_si128(a, _mm_unpackhi_epi64(a,a));
1028 return (pfirst(tmp) != 0) || (pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1)) != 0);
1035template<> EIGEN_STRONG_INLINE
float predux_mul<Packet4f>(
const Packet4f& a)
1037 Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a));
1038 return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
1040template<> EIGEN_STRONG_INLINE
double predux_mul<Packet2d>(
const Packet2d& a)
1042 return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a)));
1044template<> EIGEN_STRONG_INLINE
int predux_mul<Packet4i>(
const Packet4i& a)
1049 EIGEN_ALIGN16
int aux[4];
1051 return (aux[0] * aux[1]) * (aux[2] * aux[3]);
1054template<> EIGEN_STRONG_INLINE
bool predux_mul<Packet16b>(
const Packet16b& a) {
1055 Packet4i tmp = _mm_and_si128(a, _mm_unpackhi_epi64(a,a));
1056 return ((pfirst<Packet4i>(tmp) == 0x01010101) &&
1057 (pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1)) == 0x01010101));
1061template<> EIGEN_STRONG_INLINE
float predux_min<Packet4f>(
const Packet4f& a)
1063 Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a));
1064 return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
1066template<> EIGEN_STRONG_INLINE
double predux_min<Packet2d>(
const Packet2d& a)
1068 return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a)));
1070template<> EIGEN_STRONG_INLINE
int predux_min<Packet4i>(
const Packet4i& a)
1072#ifdef EIGEN_VECTORIZE_SSE4_1
1073 Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
1074 return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
1078 EIGEN_ALIGN16
int aux[4];
1080 int aux0 = aux[0]<aux[1] ? aux[0] : aux[1];
1081 int aux2 = aux[2]<aux[3] ? aux[2] : aux[3];
1082 return aux0<aux2 ? aux0 : aux2;
1087template<> EIGEN_STRONG_INLINE
float predux_max<Packet4f>(
const Packet4f& a)
1089 Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a));
1090 return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
1092template<> EIGEN_STRONG_INLINE
double predux_max<Packet2d>(
const Packet2d& a)
1094 return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a)));
1096template<> EIGEN_STRONG_INLINE
int predux_max<Packet4i>(
const Packet4i& a)
1098#ifdef EIGEN_VECTORIZE_SSE4_1
1099 Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
1100 return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
1104 EIGEN_ALIGN16
int aux[4];
1106 int aux0 = aux[0]>aux[1] ? aux[0] : aux[1];
1107 int aux2 = aux[2]>aux[3] ? aux[2] : aux[3];
1108 return aux0>aux2 ? aux0 : aux2;
1118template<> EIGEN_STRONG_INLINE
bool predux_any(
const Packet4f& x)
1120 return _mm_movemask_ps(x) != 0x0;
1123EIGEN_DEVICE_FUNC
inline void
1124ptranspose(PacketBlock<Packet4f,4>& kernel) {
1125 _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]);
1128EIGEN_DEVICE_FUNC
inline void
1129ptranspose(PacketBlock<Packet2d,2>& kernel) {
1130 __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
1131 kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
1132 kernel.packet[1] = tmp;
1135EIGEN_DEVICE_FUNC
inline void
1136ptranspose(PacketBlock<Packet4i,4>& kernel) {
1137 __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]);
1138 __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]);
1139 __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]);
1140 __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]);
1142 kernel.packet[0] = _mm_unpacklo_epi64(T0, T1);
1143 kernel.packet[1] = _mm_unpackhi_epi64(T0, T1);
1144 kernel.packet[2] = _mm_unpacklo_epi64(T2, T3);
1145 kernel.packet[3] = _mm_unpackhi_epi64(T2, T3);
1148EIGEN_DEVICE_FUNC
inline void
1149ptranspose(PacketBlock<Packet16b,4>& kernel) {
1150 __m128i T0 = _mm_unpacklo_epi8(kernel.packet[0], kernel.packet[1]);
1151 __m128i T1 = _mm_unpackhi_epi8(kernel.packet[0], kernel.packet[1]);
1152 __m128i T2 = _mm_unpacklo_epi8(kernel.packet[2], kernel.packet[3]);
1153 __m128i T3 = _mm_unpackhi_epi8(kernel.packet[2], kernel.packet[3]);
1154 kernel.packet[0] = _mm_unpacklo_epi16(T0, T2);
1155 kernel.packet[1] = _mm_unpackhi_epi16(T0, T2);
1156 kernel.packet[2] = _mm_unpacklo_epi16(T1, T3);
1157 kernel.packet[3] = _mm_unpackhi_epi16(T1, T3);
1160EIGEN_DEVICE_FUNC
inline void
1161ptranspose(PacketBlock<Packet16b,16>& kernel) {
1173 __m128i t0 = _mm_unpacklo_epi8(kernel.packet[0], kernel.packet[1]);
1174 __m128i t1 = _mm_unpackhi_epi8(kernel.packet[0], kernel.packet[1]);
1175 __m128i t2 = _mm_unpacklo_epi8(kernel.packet[2], kernel.packet[3]);
1176 __m128i t3 = _mm_unpackhi_epi8(kernel.packet[2], kernel.packet[3]);
1177 __m128i t4 = _mm_unpacklo_epi8(kernel.packet[4], kernel.packet[5]);
1178 __m128i t5 = _mm_unpackhi_epi8(kernel.packet[4], kernel.packet[5]);
1179 __m128i t6 = _mm_unpacklo_epi8(kernel.packet[6], kernel.packet[7]);
1180 __m128i t7 = _mm_unpackhi_epi8(kernel.packet[6], kernel.packet[7]);
1181 __m128i t8 = _mm_unpacklo_epi8(kernel.packet[8], kernel.packet[9]);
1182 __m128i t9 = _mm_unpackhi_epi8(kernel.packet[8], kernel.packet[9]);
1183 __m128i ta = _mm_unpacklo_epi8(kernel.packet[10], kernel.packet[11]);
1184 __m128i tb = _mm_unpackhi_epi8(kernel.packet[10], kernel.packet[11]);
1185 __m128i tc = _mm_unpacklo_epi8(kernel.packet[12], kernel.packet[13]);
1186 __m128i td = _mm_unpackhi_epi8(kernel.packet[12], kernel.packet[13]);
1187 __m128i te = _mm_unpacklo_epi8(kernel.packet[14], kernel.packet[15]);
1188 __m128i tf = _mm_unpackhi_epi8(kernel.packet[14], kernel.packet[15]);
1190 __m128i s0 = _mm_unpacklo_epi16(t0, t2);
1191 __m128i s1 = _mm_unpackhi_epi16(t0, t2);
1192 __m128i s2 = _mm_unpacklo_epi16(t1, t3);
1193 __m128i s3 = _mm_unpackhi_epi16(t1, t3);
1194 __m128i s4 = _mm_unpacklo_epi16(t4, t6);
1195 __m128i s5 = _mm_unpackhi_epi16(t4, t6);
1196 __m128i s6 = _mm_unpacklo_epi16(t5, t7);
1197 __m128i s7 = _mm_unpackhi_epi16(t5, t7);
1198 __m128i s8 = _mm_unpacklo_epi16(t8, ta);
1199 __m128i s9 = _mm_unpackhi_epi16(t8, ta);
1200 __m128i sa = _mm_unpacklo_epi16(t9, tb);
1201 __m128i sb = _mm_unpackhi_epi16(t9, tb);
1202 __m128i sc = _mm_unpacklo_epi16(tc, te);
1203 __m128i sd = _mm_unpackhi_epi16(tc, te);
1204 __m128i se = _mm_unpacklo_epi16(td, tf);
1205 __m128i sf = _mm_unpackhi_epi16(td, tf);
1207 __m128i u0 = _mm_unpacklo_epi32(s0, s4);
1208 __m128i u1 = _mm_unpackhi_epi32(s0, s4);
1209 __m128i u2 = _mm_unpacklo_epi32(s1, s5);
1210 __m128i u3 = _mm_unpackhi_epi32(s1, s5);
1211 __m128i u4 = _mm_unpacklo_epi32(s2, s6);
1212 __m128i u5 = _mm_unpackhi_epi32(s2, s6);
1213 __m128i u6 = _mm_unpacklo_epi32(s3, s7);
1214 __m128i u7 = _mm_unpackhi_epi32(s3, s7);
1215 __m128i u8 = _mm_unpacklo_epi32(s8, sc);
1216 __m128i u9 = _mm_unpackhi_epi32(s8, sc);
1217 __m128i ua = _mm_unpacklo_epi32(s9, sd);
1218 __m128i ub = _mm_unpackhi_epi32(s9, sd);
1219 __m128i uc = _mm_unpacklo_epi32(sa, se);
1220 __m128i ud = _mm_unpackhi_epi32(sa, se);
1221 __m128i ue = _mm_unpacklo_epi32(sb, sf);
1222 __m128i uf = _mm_unpackhi_epi32(sb, sf);
1224 kernel.packet[0] = _mm_unpacklo_epi64(u0, u8);
1225 kernel.packet[1] = _mm_unpackhi_epi64(u0, u8);
1226 kernel.packet[2] = _mm_unpacklo_epi64(u1, u9);
1227 kernel.packet[3] = _mm_unpackhi_epi64(u1, u9);
1228 kernel.packet[4] = _mm_unpacklo_epi64(u2, ua);
1229 kernel.packet[5] = _mm_unpackhi_epi64(u2, ua);
1230 kernel.packet[6] = _mm_unpacklo_epi64(u3, ub);
1231 kernel.packet[7] = _mm_unpackhi_epi64(u3, ub);
1232 kernel.packet[8] = _mm_unpacklo_epi64(u4, uc);
1233 kernel.packet[9] = _mm_unpackhi_epi64(u4, uc);
1234 kernel.packet[10] = _mm_unpacklo_epi64(u5, ud);
1235 kernel.packet[11] = _mm_unpackhi_epi64(u5, ud);
1236 kernel.packet[12] = _mm_unpacklo_epi64(u6, ue);
1237 kernel.packet[13] = _mm_unpackhi_epi64(u6, ue);
1238 kernel.packet[14] = _mm_unpacklo_epi64(u7, uf);
1239 kernel.packet[15] = _mm_unpackhi_epi64(u7, uf);
1242template<> EIGEN_STRONG_INLINE Packet4i pblend(
const Selector<4>& ifPacket,
const Packet4i& thenPacket,
const Packet4i& elsePacket) {
1243 const __m128i zero = _mm_setzero_si128();
1244 const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
1245 __m128i false_mask = _mm_cmpeq_epi32(select, zero);
1246#ifdef EIGEN_VECTORIZE_SSE4_1
1247 return _mm_blendv_epi8(thenPacket, elsePacket, false_mask);
1249 return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket));
1252template<> EIGEN_STRONG_INLINE Packet4f pblend(
const Selector<4>& ifPacket,
const Packet4f& thenPacket,
const Packet4f& elsePacket) {
1253 const __m128 zero = _mm_setzero_ps();
1254 const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
1255 __m128 false_mask = _mm_cmpeq_ps(select, zero);
1256#ifdef EIGEN_VECTORIZE_SSE4_1
1257 return _mm_blendv_ps(thenPacket, elsePacket, false_mask);
1259 return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket));
1262template<> EIGEN_STRONG_INLINE Packet2d pblend(
const Selector<2>& ifPacket,
const Packet2d& thenPacket,
const Packet2d& elsePacket) {
1263 const __m128d zero = _mm_setzero_pd();
1264 const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]);
1265 __m128d false_mask = _mm_cmpeq_pd(select, zero);
1266#ifdef EIGEN_VECTORIZE_SSE4_1
1267 return _mm_blendv_pd(thenPacket, elsePacket, false_mask);
1269 return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket));
1274#ifdef EIGEN_VECTORIZE_FMA
1275template<> EIGEN_STRONG_INLINE
float pmadd(
const float& a,
const float& b,
const float& c) {
1276 return ::fmaf(a,b,c);
1278template<> EIGEN_STRONG_INLINE
double pmadd(
const double& a,
const double& b,
const double& c) {
1279 return ::fma(a,b,c);
1294template<>
struct is_arithmetic<Packet4h> {
enum { value =
true }; };
1297struct packet_traits<
Eigen::half> : default_packet_traits {
1298 typedef Packet4h type;
1300 typedef Packet4h half;
1303 AlignedOnScalar = 1,
1326template<>
struct unpacket_traits<Packet4h> {
typedef Eigen::half type;
enum {size=4, alignment=
Aligned16, vectorizable=
true, masked_load_available=
false, masked_store_available=
false};
typedef Packet4h half; };
1328template<> EIGEN_STRONG_INLINE Packet4h pset1<Packet4h>(
const Eigen::half& from) {
1330 result.x = _mm_set1_pi16(from.x);
1334template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet4h>(
const Packet4h& from) {
1335 return half_impl::raw_uint16_to_half(
static_cast<unsigned short>(_mm_cvtsi64_si32(from.x)));
1338template<> EIGEN_STRONG_INLINE Packet4h pconj(
const Packet4h& a) {
return a; }
1340template<> EIGEN_STRONG_INLINE Packet4h padd<Packet4h>(
const Packet4h& a,
const Packet4h& b) {
1341 __int64_t a64 = _mm_cvtm64_si64(a.x);
1342 __int64_t b64 = _mm_cvtm64_si64(b.x);
1346 Eigen::half ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64));
1347 Eigen::half hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64));
1349 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 16));
1350 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 16));
1352 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 32));
1353 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 32));
1355 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 48));
1356 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 48));
1359 result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
1363template<> EIGEN_STRONG_INLINE Packet4h psub<Packet4h>(
const Packet4h& a,
const Packet4h& b) {
1364 __int64_t a64 = _mm_cvtm64_si64(a.x);
1365 __int64_t b64 = _mm_cvtm64_si64(b.x);
1369 Eigen::half ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64));
1370 Eigen::half hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64));
1372 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 16));
1373 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 16));
1375 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 32));
1376 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 32));
1378 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 48));
1379 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 48));
1382 result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
1386template<> EIGEN_STRONG_INLINE Packet4h pmul<Packet4h>(
const Packet4h& a,
const Packet4h& b) {
1387 __int64_t a64 = _mm_cvtm64_si64(a.x);
1388 __int64_t b64 = _mm_cvtm64_si64(b.x);
1392 Eigen::half ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64));
1393 Eigen::half hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64));
1395 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 16));
1396 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 16));
1398 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 32));
1399 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 32));
1401 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 48));
1402 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 48));
1405 result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
1409template<> EIGEN_STRONG_INLINE Packet4h pdiv<Packet4h>(
const Packet4h& a,
const Packet4h& b) {
1410 __int64_t a64 = _mm_cvtm64_si64(a.x);
1411 __int64_t b64 = _mm_cvtm64_si64(b.x);
1415 Eigen::half ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64));
1416 Eigen::half hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64));
1418 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 16));
1419 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 16));
1421 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 32));
1422 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 32));
1424 ha = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(a64 >> 48));
1425 hb = half_impl::raw_uint16_to_half(
static_cast<unsigned short>(b64 >> 48));
1428 result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
1432template<> EIGEN_STRONG_INLINE Packet4h pload<Packet4h>(
const Eigen::half* from) {
1434 result.x = _mm_cvtsi64_m64(*
reinterpret_cast<const __int64_t*
>(from));
1438template<> EIGEN_STRONG_INLINE Packet4h ploadu<Packet4h>(
const Eigen::half* from) {
1440 result.x = _mm_cvtsi64_m64(*
reinterpret_cast<const __int64_t*
>(from));
1444template<> EIGEN_STRONG_INLINE
void pstore<Eigen::half>(Eigen::half* to,
const Packet4h& from) {
1445 __int64_t r = _mm_cvtm64_si64(from.x);
1446 *(
reinterpret_cast<__int64_t*
>(to)) = r;
1449template<> EIGEN_STRONG_INLINE
void pstoreu<Eigen::half>(Eigen::half* to,
const Packet4h& from) {
1450 __int64_t r = _mm_cvtm64_si64(from.x);
1451 *(
reinterpret_cast<__int64_t*
>(to)) = r;
1454template<> EIGEN_STRONG_INLINE Packet4h
1455ploadquad<Packet4h>(
const Eigen::half* from) {
1456 return pset1<Packet4h>(*from);
1459template<> EIGEN_STRONG_INLINE Packet4h pgather<Eigen::half, Packet4h>(
const Eigen::half* from,
Index stride)
1462 result.x = _mm_set_pi16(from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
1466template<> EIGEN_STRONG_INLINE
void pscatter<Eigen::half, Packet4h>(Eigen::half* to,
const Packet4h& from,
Index stride)
1468 __int64_t a = _mm_cvtm64_si64(from.x);
1469 to[stride*0].x =
static_cast<unsigned short>(a);
1470 to[stride*1].x =
static_cast<unsigned short>(a >> 16);
1471 to[stride*2].x =
static_cast<unsigned short>(a >> 32);
1472 to[stride*3].x =
static_cast<unsigned short>(a >> 48);
1475EIGEN_STRONG_INLINE
void
1476ptranspose(PacketBlock<Packet4h,4>& kernel) {
1477 __m64 T0 = _mm_unpacklo_pi16(kernel.packet[0].x, kernel.packet[1].x);
1478 __m64 T1 = _mm_unpacklo_pi16(kernel.packet[2].x, kernel.packet[3].x);
1479 __m64 T2 = _mm_unpackhi_pi16(kernel.packet[0].x, kernel.packet[1].x);
1480 __m64 T3 = _mm_unpackhi_pi16(kernel.packet[2].x, kernel.packet[3].x);
1482 kernel.packet[0].x = _mm_unpacklo_pi32(T0, T1);
1483 kernel.packet[1].x = _mm_unpackhi_pi32(T0, T1);
1484 kernel.packet[2].x = _mm_unpacklo_pi32(T2, T3);
1485 kernel.packet[3].x = _mm_unpackhi_pi32(T2, T3);
1495#if EIGEN_COMP_PGI && EIGEN_COMP_PGI < 1900
1497static inline __m128 _mm_castpd_ps (__m128d x) {
return reinterpret_cast<__m128&
>(x); }
1498static inline __m128i _mm_castpd_si128(__m128d x) {
return reinterpret_cast<__m128i&
>(x); }
1499static inline __m128d _mm_castps_pd (__m128 x) {
return reinterpret_cast<__m128d&
>(x); }
1500static inline __m128i _mm_castps_si128(__m128 x) {
return reinterpret_cast<__m128i&
>(x); }
1501static inline __m128 _mm_castsi128_ps(__m128i x) {
return reinterpret_cast<__m128&
>(x); }
1502static inline __m128d _mm_castsi128_pd(__m128i x) {
return reinterpret_cast<__m128d&
>(x); }
@ Aligned16
Definition: Constants.h:235
Namespace containing all symbols from the Eigen library.
Definition: Core:141
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index
The Index type as used for the API.
Definition: Meta.h:74