12 #ifndef EIGEN_PACKET_MATH_NEON_H 13 #define EIGEN_PACKET_MATH_NEON_H 19 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 20 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 23 #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD 24 #define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 27 #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD 28 #define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD 31 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 33 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32 35 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 39 typedef float32x2_t Packet2f;
40 typedef float32x4_t Packet4f;
41 typedef int32x4_t Packet4i;
42 typedef int32x2_t Packet2i;
43 typedef uint32x4_t Packet4ui;
45 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ 46 const Packet4f p4f_##NAME = pset1<Packet4f>(X) 48 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ 49 const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X)) 51 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ 52 const Packet4i p4i_##NAME = pset1<Packet4i>(X) 56 #if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC 57 #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR); 59 #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR) 60 #elif !EIGEN_ARCH_ARM64 61 #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" ); 64 #define EIGEN_ARM_PREFETCH(ADDR) 67 template<>
struct packet_traits<float> : default_packet_traits
69 typedef Packet4f type;
70 typedef Packet4f half;
86 template<>
struct packet_traits<int> : default_packet_traits
88 typedef Packet4i type;
89 typedef Packet4i half;
99 #if EIGEN_GNUC_AT_MOST(4,4) && !EIGEN_COMP_LLVM 101 EIGEN_STRONG_INLINE float32x4_t vld1q_f32(
const float* x) { return ::vld1q_f32((
const float32_t*)x); }
102 EIGEN_STRONG_INLINE float32x2_t vld1_f32 (
const float* x) { return ::vld1_f32 ((
const float32_t*)x); }
103 EIGEN_STRONG_INLINE float32x2_t vld1_dup_f32 (
const float* x) { return ::vld1_dup_f32 ((
const float32_t*)x); }
104 EIGEN_STRONG_INLINE
void vst1q_f32(
float* to, float32x4_t from) { ::vst1q_f32((float32_t*)to,from); }
105 EIGEN_STRONG_INLINE
void vst1_f32 (
float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); }
108 template<>
struct unpacket_traits<Packet4f> {
typedef float type;
enum {size=4, alignment=
Aligned16};
typedef Packet4f half; };
109 template<>
struct unpacket_traits<Packet4i> {
typedef int type;
enum {size=4, alignment=
Aligned16};
typedef Packet4i half; };
111 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(
const float& from) {
return vdupq_n_f32(from); }
112 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(
const int& from) {
return vdupq_n_s32(from); }
114 template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(
const float& a)
116 const float32_t f[] = {0, 1, 2, 3};
117 Packet4f countdown = vld1q_f32(f);
118 return vaddq_f32(pset1<Packet4f>(a), countdown);
120 template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(
const int& a)
122 const int32_t i[] = {0, 1, 2, 3};
123 Packet4i countdown = vld1q_s32(i);
124 return vaddq_s32(pset1<Packet4i>(a), countdown);
127 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return vaddq_f32(a,b); }
128 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return vaddq_s32(a,b); }
130 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return vsubq_f32(a,b); }
131 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return vsubq_s32(a,b); }
133 template<> EIGEN_STRONG_INLINE Packet4f pnegate(
const Packet4f& a) {
return vnegq_f32(a); }
134 template<> EIGEN_STRONG_INLINE Packet4i pnegate(
const Packet4i& a) {
return vnegq_s32(a); }
136 template<> EIGEN_STRONG_INLINE Packet4f pconj(
const Packet4f& a) {
return a; }
137 template<> EIGEN_STRONG_INLINE Packet4i pconj(
const Packet4i& a) {
return a; }
139 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return vmulq_f32(a,b); }
140 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return vmulq_s32(a,b); }
142 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(
const Packet4f& a,
const Packet4f& b)
145 return vdivq_f32(a,b);
147 Packet4f inv, restep, div;
154 inv = vrecpeq_f32(b);
158 restep = vrecpsq_f32(b, inv);
159 inv = vmulq_f32(restep, inv);
162 div = vmulq_f32(a, inv);
168 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(
const Packet4i& ,
const Packet4i& )
169 { eigen_assert(
false &&
"packet integer division are not supported by NEON");
170 return pset1<Packet4i>(0);
177 #if (defined __ARM_FEATURE_FMA) && !(EIGEN_COMP_CLANG && EIGEN_ARCH_ARM) 184 template<> EIGEN_STRONG_INLINE Packet4f pmadd(
const Packet4f& a,
const Packet4f& b,
const Packet4f& c) {
return vfmaq_f32(c,a,b); }
186 template<> EIGEN_STRONG_INLINE Packet4f pmadd(
const Packet4f& a,
const Packet4f& b,
const Packet4f& c) {
187 #if EIGEN_COMP_CLANG && EIGEN_ARCH_ARM 197 "vmla.f32 %q[r], %q[a], %q[b]" 204 return vmlaq_f32(c,a,b);
210 template<> EIGEN_STRONG_INLINE Packet4i pmadd(
const Packet4i& a,
const Packet4i& b,
const Packet4i& c) {
return vmlaq_s32(c,a,b); }
212 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return vminq_f32(a,b); }
213 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return vminq_s32(a,b); }
215 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return vmaxq_f32(a,b); }
216 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return vmaxq_s32(a,b); }
219 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(
const Packet4f& a,
const Packet4f& b)
221 return vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
223 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return vandq_s32(a,b); }
225 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(
const Packet4f& a,
const Packet4f& b)
227 return vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
229 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return vorrq_s32(a,b); }
231 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(
const Packet4f& a,
const Packet4f& b)
233 return vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
235 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return veorq_s32(a,b); }
237 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(
const Packet4f& a,
const Packet4f& b)
239 return vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
241 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return vbicq_s32(a,b); }
243 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(
const float* from) { EIGEN_DEBUG_ALIGNED_LOAD
return vld1q_f32(from); }
244 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(
const int* from) { EIGEN_DEBUG_ALIGNED_LOAD
return vld1q_s32(from); }
246 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(
const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD
return vld1q_f32(from); }
247 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(
const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD
return vld1q_s32(from); }
249 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(
const float* from)
252 lo = vld1_dup_f32(from);
253 hi = vld1_dup_f32(from+1);
254 return vcombine_f32(lo, hi);
256 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(
const int* from)
259 lo = vld1_dup_s32(from);
260 hi = vld1_dup_s32(from+1);
261 return vcombine_s32(lo, hi);
264 template<> EIGEN_STRONG_INLINE
void pstore<float>(
float* to,
const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); }
265 template<> EIGEN_STRONG_INLINE
void pstore<int>(
int* to,
const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); }
267 template<> EIGEN_STRONG_INLINE
void pstoreu<float>(
float* to,
const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
268 template<> EIGEN_STRONG_INLINE
void pstoreu<int>(
int* to,
const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
270 template<> EIGEN_DEVICE_FUNC
inline Packet4f pgather<float, Packet4f>(
const float* from,
Index stride)
272 Packet4f res = pset1<Packet4f>(0.f);
273 res = vsetq_lane_f32(from[0*stride], res, 0);
274 res = vsetq_lane_f32(from[1*stride], res, 1);
275 res = vsetq_lane_f32(from[2*stride], res, 2);
276 res = vsetq_lane_f32(from[3*stride], res, 3);
279 template<> EIGEN_DEVICE_FUNC
inline Packet4i pgather<int, Packet4i>(
const int* from,
Index stride)
281 Packet4i res = pset1<Packet4i>(0);
282 res = vsetq_lane_s32(from[0*stride], res, 0);
283 res = vsetq_lane_s32(from[1*stride], res, 1);
284 res = vsetq_lane_s32(from[2*stride], res, 2);
285 res = vsetq_lane_s32(from[3*stride], res, 3);
289 template<> EIGEN_DEVICE_FUNC
inline void pscatter<float, Packet4f>(
float* to,
const Packet4f& from,
Index stride)
291 to[stride*0] = vgetq_lane_f32(from, 0);
292 to[stride*1] = vgetq_lane_f32(from, 1);
293 to[stride*2] = vgetq_lane_f32(from, 2);
294 to[stride*3] = vgetq_lane_f32(from, 3);
296 template<> EIGEN_DEVICE_FUNC
inline void pscatter<int, Packet4i>(
int* to,
const Packet4i& from,
Index stride)
298 to[stride*0] = vgetq_lane_s32(from, 0);
299 to[stride*1] = vgetq_lane_s32(from, 1);
300 to[stride*2] = vgetq_lane_s32(from, 2);
301 to[stride*3] = vgetq_lane_s32(from, 3);
304 template<> EIGEN_STRONG_INLINE
void prefetch<float>(
const float* addr) { EIGEN_ARM_PREFETCH(addr); }
305 template<> EIGEN_STRONG_INLINE
void prefetch<int>(
const int* addr) { EIGEN_ARM_PREFETCH(addr); }
308 template<> EIGEN_STRONG_INLINE
float pfirst<Packet4f>(
const Packet4f& a) {
float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a);
return x[0]; }
309 template<> EIGEN_STRONG_INLINE
int pfirst<Packet4i>(
const Packet4i& a) {
int EIGEN_ALIGN16 x[4]; vst1q_s32(x, a);
return x[0]; }
311 template<> EIGEN_STRONG_INLINE Packet4f preverse(
const Packet4f& a) {
312 float32x2_t a_lo, a_hi;
315 a_r64 = vrev64q_f32(a);
316 a_lo = vget_low_f32(a_r64);
317 a_hi = vget_high_f32(a_r64);
318 return vcombine_f32(a_hi, a_lo);
320 template<> EIGEN_STRONG_INLINE Packet4i preverse(
const Packet4i& a) {
321 int32x2_t a_lo, a_hi;
324 a_r64 = vrev64q_s32(a);
325 a_lo = vget_low_s32(a_r64);
326 a_hi = vget_high_s32(a_r64);
327 return vcombine_s32(a_hi, a_lo);
330 template<> EIGEN_STRONG_INLINE Packet4f pabs(
const Packet4f& a) {
return vabsq_f32(a); }
331 template<> EIGEN_STRONG_INLINE Packet4i pabs(
const Packet4i& a) {
return vabsq_s32(a); }
333 template<> EIGEN_STRONG_INLINE
float predux<Packet4f>(
const Packet4f& a)
335 float32x2_t a_lo, a_hi, sum;
337 a_lo = vget_low_f32(a);
338 a_hi = vget_high_f32(a);
339 sum = vpadd_f32(a_lo, a_hi);
340 sum = vpadd_f32(sum, sum);
341 return vget_lane_f32(sum, 0);
344 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(
const Packet4f* vecs)
346 float32x4x2_t vtrn1, vtrn2, res1, res2;
347 Packet4f sum1, sum2, sum;
351 vtrn1 = vzipq_f32(vecs[0], vecs[2]);
352 vtrn2 = vzipq_f32(vecs[1], vecs[3]);
353 res1 = vzipq_f32(vtrn1.val[0], vtrn2.val[0]);
354 res2 = vzipq_f32(vtrn1.val[1], vtrn2.val[1]);
357 sum1 = vaddq_f32(res1.val[0], res1.val[1]);
358 sum2 = vaddq_f32(res2.val[0], res2.val[1]);
359 sum = vaddq_f32(sum1, sum2);
364 template<> EIGEN_STRONG_INLINE
int predux<Packet4i>(
const Packet4i& a)
366 int32x2_t a_lo, a_hi, sum;
368 a_lo = vget_low_s32(a);
369 a_hi = vget_high_s32(a);
370 sum = vpadd_s32(a_lo, a_hi);
371 sum = vpadd_s32(sum, sum);
372 return vget_lane_s32(sum, 0);
375 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(
const Packet4i* vecs)
377 int32x4x2_t vtrn1, vtrn2, res1, res2;
378 Packet4i sum1, sum2, sum;
382 vtrn1 = vzipq_s32(vecs[0], vecs[2]);
383 vtrn2 = vzipq_s32(vecs[1], vecs[3]);
384 res1 = vzipq_s32(vtrn1.val[0], vtrn2.val[0]);
385 res2 = vzipq_s32(vtrn1.val[1], vtrn2.val[1]);
388 sum1 = vaddq_s32(res1.val[0], res1.val[1]);
389 sum2 = vaddq_s32(res2.val[0], res2.val[1]);
390 sum = vaddq_s32(sum1, sum2);
397 template<> EIGEN_STRONG_INLINE
float predux_mul<Packet4f>(
const Packet4f& a)
399 float32x2_t a_lo, a_hi, prod;
402 a_lo = vget_low_f32(a);
403 a_hi = vget_high_f32(a);
405 prod = vmul_f32(a_lo, a_hi);
407 prod = vmul_f32(prod, vrev64_f32(prod));
409 return vget_lane_f32(prod, 0);
411 template<> EIGEN_STRONG_INLINE
int predux_mul<Packet4i>(
const Packet4i& a)
413 int32x2_t a_lo, a_hi, prod;
416 a_lo = vget_low_s32(a);
417 a_hi = vget_high_s32(a);
419 prod = vmul_s32(a_lo, a_hi);
421 prod = vmul_s32(prod, vrev64_s32(prod));
423 return vget_lane_s32(prod, 0);
427 template<> EIGEN_STRONG_INLINE
float predux_min<Packet4f>(
const Packet4f& a)
429 float32x2_t a_lo, a_hi, min;
431 a_lo = vget_low_f32(a);
432 a_hi = vget_high_f32(a);
433 min = vpmin_f32(a_lo, a_hi);
434 min = vpmin_f32(min, min);
436 return vget_lane_f32(min, 0);
439 template<> EIGEN_STRONG_INLINE
int predux_min<Packet4i>(
const Packet4i& a)
441 int32x2_t a_lo, a_hi, min;
443 a_lo = vget_low_s32(a);
444 a_hi = vget_high_s32(a);
445 min = vpmin_s32(a_lo, a_hi);
446 min = vpmin_s32(min, min);
448 return vget_lane_s32(min, 0);
452 template<> EIGEN_STRONG_INLINE
float predux_max<Packet4f>(
const Packet4f& a)
454 float32x2_t a_lo, a_hi, max;
456 a_lo = vget_low_f32(a);
457 a_hi = vget_high_f32(a);
458 max = vpmax_f32(a_lo, a_hi);
459 max = vpmax_f32(max, max);
461 return vget_lane_f32(max, 0);
464 template<> EIGEN_STRONG_INLINE
int predux_max<Packet4i>(
const Packet4i& a)
466 int32x2_t a_lo, a_hi, max;
468 a_lo = vget_low_s32(a);
469 a_hi = vget_high_s32(a);
470 max = vpmax_s32(a_lo, a_hi);
471 max = vpmax_s32(max, max);
473 return vget_lane_s32(max, 0);
478 #define PALIGN_NEON(Offset,Type,Command) \ 480 struct palign_impl<Offset,Type>\ 482 EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\ 485 first = Command(first, second, Offset);\ 489 PALIGN_NEON(0,Packet4f,vextq_f32)
490 PALIGN_NEON(1,Packet4f,vextq_f32)
491 PALIGN_NEON(2,Packet4f,vextq_f32)
492 PALIGN_NEON(3,Packet4f,vextq_f32)
493 PALIGN_NEON(0,Packet4i,vextq_s32)
494 PALIGN_NEON(1,Packet4i,vextq_s32)
495 PALIGN_NEON(2,Packet4i,vextq_s32)
496 PALIGN_NEON(3,Packet4i,vextq_s32)
500 EIGEN_DEVICE_FUNC
inline void 501 ptranspose(PacketBlock<Packet4f,4>& kernel) {
502 float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]);
503 float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]);
505 kernel.packet[0] = vcombine_f32(vget_low_f32(tmp1.val[0]), vget_low_f32(tmp2.val[0]));
506 kernel.packet[1] = vcombine_f32(vget_high_f32(tmp1.val[0]), vget_high_f32(tmp2.val[0]));
507 kernel.packet[2] = vcombine_f32(vget_low_f32(tmp1.val[1]), vget_low_f32(tmp2.val[1]));
508 kernel.packet[3] = vcombine_f32(vget_high_f32(tmp1.val[1]), vget_high_f32(tmp2.val[1]));
511 EIGEN_DEVICE_FUNC
inline void 512 ptranspose(PacketBlock<Packet4i,4>& kernel) {
513 int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]);
514 int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]);
515 kernel.packet[0] = vcombine_s32(vget_low_s32(tmp1.val[0]), vget_low_s32(tmp2.val[0]));
516 kernel.packet[1] = vcombine_s32(vget_high_s32(tmp1.val[0]), vget_high_s32(tmp2.val[0]));
517 kernel.packet[2] = vcombine_s32(vget_low_s32(tmp1.val[1]), vget_low_s32(tmp2.val[1]));
518 kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1]));
525 #ifdef __apple_build_version__ 529 #define EIGEN_APPLE_DOUBLE_NEON_BUG (__apple_build_version__ < 6010000) 531 #define EIGEN_APPLE_DOUBLE_NEON_BUG 0 534 #if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG 540 template <
typename T>
541 uint64x2_t vreinterpretq_u64_f64(T a)
543 return (uint64x2_t) a;
546 template <
typename T>
547 float64x2_t vreinterpretq_f64_u64(T a)
549 return (float64x2_t) a;
552 typedef float64x2_t Packet2d;
553 typedef float64x1_t Packet1d;
555 template<>
struct packet_traits<double> : default_packet_traits
557 typedef Packet2d type;
558 typedef Packet2d half;
575 template<>
struct unpacket_traits<Packet2d> {
typedef double type;
enum {size=2, alignment=
Aligned16};
typedef Packet2d half; };
577 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(
const double& from) {
return vdupq_n_f64(from); }
579 template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(
const double& a)
581 const double countdown_raw[] = {0.0,1.0};
582 const Packet2d countdown = vld1q_f64(countdown_raw);
583 return vaddq_f64(pset1<Packet2d>(a), countdown);
585 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return vaddq_f64(a,b); }
587 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return vsubq_f64(a,b); }
589 template<> EIGEN_STRONG_INLINE Packet2d pnegate(
const Packet2d& a) {
return vnegq_f64(a); }
591 template<> EIGEN_STRONG_INLINE Packet2d pconj(
const Packet2d& a) {
return a; }
593 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return vmulq_f64(a,b); }
595 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return vdivq_f64(a,b); }
597 #ifdef __ARM_FEATURE_FMA 599 template<> EIGEN_STRONG_INLINE Packet2d pmadd(
const Packet2d& a,
const Packet2d& b,
const Packet2d& c) {
return vfmaq_f64(c,a,b); }
601 template<> EIGEN_STRONG_INLINE Packet2d pmadd(
const Packet2d& a,
const Packet2d& b,
const Packet2d& c) {
return vmlaq_f64(c,a,b); }
604 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return vminq_f64(a,b); }
606 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return vmaxq_f64(a,b); }
609 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(
const Packet2d& a,
const Packet2d& b)
611 return vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
614 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(
const Packet2d& a,
const Packet2d& b)
616 return vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
619 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(
const Packet2d& a,
const Packet2d& b)
621 return vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
624 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(
const Packet2d& a,
const Packet2d& b)
626 return vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
629 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(
const double* from) { EIGEN_DEBUG_ALIGNED_LOAD
return vld1q_f64(from); }
631 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(
const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD
return vld1q_f64(from); }
633 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(
const double* from)
635 return vld1q_dup_f64(from);
637 template<> EIGEN_STRONG_INLINE
void pstore<double>(
double* to,
const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f64(to, from); }
639 template<> EIGEN_STRONG_INLINE
void pstoreu<double>(
double* to,
const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f64(to, from); }
641 template<> EIGEN_DEVICE_FUNC
inline Packet2d pgather<double, Packet2d>(
const double* from,
Index stride)
643 Packet2d res = pset1<Packet2d>(0.0);
644 res = vsetq_lane_f64(from[0*stride], res, 0);
645 res = vsetq_lane_f64(from[1*stride], res, 1);
648 template<> EIGEN_DEVICE_FUNC
inline void pscatter<double, Packet2d>(
double* to,
const Packet2d& from,
Index stride)
650 to[stride*0] = vgetq_lane_f64(from, 0);
651 to[stride*1] = vgetq_lane_f64(from, 1);
653 template<> EIGEN_STRONG_INLINE
void prefetch<double>(
const double* addr) { EIGEN_ARM_PREFETCH(addr); }
656 template<> EIGEN_STRONG_INLINE
double pfirst<Packet2d>(
const Packet2d& a) {
return vgetq_lane_f64(a, 0); }
658 template<> EIGEN_STRONG_INLINE Packet2d preverse(
const Packet2d& a) {
return vcombine_f64(vget_high_f64(a), vget_low_f64(a)); }
660 template<> EIGEN_STRONG_INLINE Packet2d pabs(
const Packet2d& a) {
return vabsq_f64(a); }
662 #if EIGEN_COMP_CLANG && defined(__apple_build_version__) 664 template<> EIGEN_STRONG_INLINE
double predux<Packet2d>(
const Packet2d& a) {
return (vget_low_f64(a) + vget_high_f64(a))[0]; }
666 template<> EIGEN_STRONG_INLINE
double predux<Packet2d>(
const Packet2d& a) {
return vget_lane_f64(vget_low_f64(a) + vget_high_f64(a), 0); }
669 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(
const Packet2d* vecs)
671 float64x2_t trn1, trn2;
675 trn1 = vzip1q_f64(vecs[0], vecs[1]);
676 trn2 = vzip2q_f64(vecs[0], vecs[1]);
679 return vaddq_f64(trn1, trn2);
683 #if EIGEN_COMP_CLANG && defined(__apple_build_version__) 684 template<> EIGEN_STRONG_INLINE
double predux_mul<Packet2d>(
const Packet2d& a) {
return (vget_low_f64(a) * vget_high_f64(a))[0]; }
686 template<> EIGEN_STRONG_INLINE
double predux_mul<Packet2d>(
const Packet2d& a) {
return vget_lane_f64(vget_low_f64(a) * vget_high_f64(a), 0); }
690 template<> EIGEN_STRONG_INLINE
double predux_min<Packet2d>(
const Packet2d& a) {
return vgetq_lane_f64(vpminq_f64(a, a), 0); }
693 template<> EIGEN_STRONG_INLINE
double predux_max<Packet2d>(
const Packet2d& a) {
return vgetq_lane_f64(vpmaxq_f64(a, a), 0); }
697 #define PALIGN_NEON(Offset,Type,Command) \ 699 struct palign_impl<Offset,Type>\ 701 EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\ 704 first = Command(first, second, Offset);\ 708 PALIGN_NEON(0,Packet2d,vextq_f64)
709 PALIGN_NEON(1,Packet2d,vextq_f64)
712 EIGEN_DEVICE_FUNC
inline void 713 ptranspose(PacketBlock<Packet2d,2>& kernel) {
714 float64x2_t trn1 = vzip1q_f64(kernel.packet[0], kernel.packet[1]);
715 float64x2_t trn2 = vzip2q_f64(kernel.packet[0], kernel.packet[1]);
717 kernel.packet[0] = trn1;
718 kernel.packet[1] = trn2;
720 #endif // EIGEN_ARCH_ARM64 726 #endif // EIGEN_PACKET_MATH_NEON_H Definition: Constants.h:230
Namespace containing all symbols from the Eigen library.
Definition: Core:287
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index
The Index type as used for the API.
Definition: Meta.h:33
Definition: Eigen_Colamd.h:50