| 1 |
//@HEADER |
| 2 |
// ************************************************************************ |
| 3 |
// |
| 4 |
// Kokkos v. 4.0 |
| 5 |
// Copyright (2022) National Technology & Engineering |
| 6 |
// Solutions of Sandia, LLC (NTESS). |
| 7 |
// |
| 8 |
// Under the terms of Contract DE-NA0003525 with NTESS, |
| 9 |
// the U.S. Government retains certain rights in this software. |
| 10 |
// |
| 11 |
// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions. |
| 12 |
// See https://kokkos.org/LICENSE for license information. |
| 13 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 14 |
// |
| 15 |
//@HEADER |
| 16 |
#ifndef KOKKOSBATCHED_VECTOR_SIMD_HPP |
| 17 |
#define KOKKOSBATCHED_VECTOR_SIMD_HPP |
| 18 |
|
| 19 |
/// \author Kyungjoo Kim ([email protected]) |
| 20 |
|
| 21 |
#include <Kokkos_Complex.hpp> |
| 22 |
#include <KokkosBatched_Vector.hpp> |
| 23 |
#include "KokkosKernels_Macros.hpp" |
| 24 |
|
| 25 |
#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__) |
| 26 |
#undef __KOKKOSBATCHED_ENABLE_AVX__ |
| 27 |
#else |
| 28 |
// compiler bug with AVX in some architectures |
| 29 |
#define __KOKKOSBATCHED_ENABLE_AVX__ |
| 30 |
#endif |
| 31 |
|
| 32 |
namespace KokkosBatched { |
| 33 |
|
| 34 |
template <typename T, int l> |
| 35 |
class Vector<SIMD<T>, l> { |
| 36 |
public: |
| 37 |
using type = Vector<SIMD<T>, l>; |
| 38 |
using value_type = T; |
| 39 |
using mag_type = typename Kokkos::ArithTraits<T>::mag_type; |
| 40 |
|
| 41 |
enum : int { vector_length = l }; |
| 42 |
|
| 43 |
typedef value_type data_type[vector_length]; |
| 44 |
|
| 45 |
KOKKOS_INLINE_FUNCTION |
| 46 |
static const char *label() { return "SIMD"; } |
| 47 |
|
| 48 |
template <typename, int> |
| 49 |
friend class Vector; |
| 50 |
|
| 51 |
private: |
| 52 |
mutable data_type _data; |
| 53 |
|
| 54 |
public: |
| 55 |
KOKKOS_INLINE_FUNCTION Vector() { |
| 56 |
// NOTE Not meant to be instantiated for CUDA |
| 57 |
KOKKOSKERNELS_FORCE_SIMD |
| 58 |
for (int i = 0; i < vector_length; ++i) _data[i] = 0; |
| 59 |
} |
| 60 |
template <typename ArgValueType> |
| 61 |
KOKKOS_INLINE_FUNCTION Vector(const ArgValueType &val) { |
| 62 |
KOKKOSKERNELS_FORCE_SIMD |
| 63 |
for (int i = 0; i < vector_length; ++i) _data[i] = val; |
| 64 |
} |
| 65 |
template <typename ArgValueType> |
| 66 |
KOKKOS_INLINE_FUNCTION Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 67 |
KOKKOSKERNELS_FORCE_SIMD |
| 68 |
for (int i = 0; i < vector_length; ++i) _data[i] = b[i]; |
| 69 |
} |
| 70 |
|
| 71 |
KOKKOS_INLINE_FUNCTION |
| 72 |
type &loadAligned(const value_type *p) { |
| 73 |
KOKKOSKERNELS_FORCE_SIMD |
| 74 |
for (int i = 0; i < vector_length; ++i) _data[i] = p[i]; |
| 75 |
return *this; |
| 76 |
} |
| 77 |
|
| 78 |
KOKKOS_INLINE_FUNCTION |
| 79 |
type &loadUnaligned(const value_type *p) { return loadAligned(p); } |
| 80 |
|
| 81 |
KOKKOS_INLINE_FUNCTION |
| 82 |
void storeAligned(value_type *p) const { |
| 83 |
KOKKOSKERNELS_FORCE_SIMD |
| 84 |
for (int i = 0; i < vector_length; ++i) p[i] = _data[i]; |
| 85 |
} |
| 86 |
|
| 87 |
KOKKOS_INLINE_FUNCTION |
| 88 |
void storeUnaligned(value_type *p) const { storeAligned(p); } |
| 89 |
|
| 90 |
KOKKOS_INLINE_FUNCTION |
| 91 |
value_type &operator[](const int &i) const { return _data[i]; } |
| 92 |
}; |
| 93 |
} // namespace KokkosBatched |
| 94 |
|
| 95 |
#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__) |
| 96 |
namespace KokkosBatched { |
| 97 |
|
| 98 |
template <> |
| 99 |
class Vector<SIMD<float>, 2> { |
| 100 |
public: |
| 101 |
using type = Vector<SIMD<float>, 2>; |
| 102 |
using value_type = float; |
| 103 |
using mag_type = float; |
| 104 |
|
| 105 |
enum : int { vector_length = 2 }; |
| 106 |
typedef float2 data_type; |
| 107 |
|
| 108 |
KOKKOS_INLINE_FUNCTION |
| 109 |
static const char *label() { return "GpuFloat2"; } |
| 110 |
|
| 111 |
template <typename, int> |
| 112 |
friend class Vector; |
| 113 |
|
| 114 |
private: |
| 115 |
mutable data_type _data; |
| 116 |
|
| 117 |
public: |
| 118 |
KOKKOS_INLINE_FUNCTION Vector() { |
| 119 |
_data.x = 0; |
| 120 |
_data.y = 0; |
| 121 |
} |
| 122 |
KOKKOS_INLINE_FUNCTION Vector(const value_type &val) { |
| 123 |
_data.x = val; |
| 124 |
_data.y = val; |
| 125 |
} |
| 126 |
KOKKOS_INLINE_FUNCTION Vector(const type &b) { |
| 127 |
_data.x = b._data.x; |
| 128 |
_data.y = b._data.y; |
| 129 |
} |
| 130 |
KOKKOS_INLINE_FUNCTION Vector(const float2 &val) { |
| 131 |
_data.x = val.x; |
| 132 |
_data.y = val.y; |
| 133 |
} |
| 134 |
|
| 135 |
template <typename ArgValueType> |
| 136 |
KOKKOS_INLINE_FUNCTION Vector(const ArgValueType &val) { |
| 137 |
_data.x = val; |
| 138 |
_data.y = val; |
| 139 |
} |
| 140 |
|
| 141 |
template <typename ArgValueType> |
| 142 |
KOKKOS_INLINE_FUNCTION Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 143 |
_data.x = b[0]; |
| 144 |
_data.y = b[1]; |
| 145 |
} |
| 146 |
|
| 147 |
KOKKOS_INLINE_FUNCTION |
| 148 |
type &operator=(const float2 &val) { |
| 149 |
_data.x = val.x; |
| 150 |
_data.y = val.y; |
| 151 |
return *this; |
| 152 |
} |
| 153 |
|
| 154 |
KOKKOS_INLINE_FUNCTION |
| 155 |
float2 float2() const { return _data; } |
| 156 |
|
| 157 |
KOKKOS_INLINE_FUNCTION |
| 158 |
type &loadAligned(const value_type *p) { |
| 159 |
_data.x = *(p); |
| 160 |
_data.y = *(p + 1); |
| 161 |
return *this; |
| 162 |
} |
| 163 |
|
| 164 |
KOKKOS_INLINE_FUNCTION |
| 165 |
type &loadUnaligned(const value_type *p) { |
| 166 |
_data.x = *(p); |
| 167 |
_data.y = *(p + 1); |
| 168 |
return *this; |
| 169 |
} |
| 170 |
|
| 171 |
KOKKOS_INLINE_FUNCTION |
| 172 |
void storeAligned(value_type *p) const { |
| 173 |
*(p) = _data.x; |
| 174 |
*(p + 1) = _data.y; |
| 175 |
} |
| 176 |
|
| 177 |
KOKKOS_INLINE_FUNCTION |
| 178 |
void storeUnaligned(value_type *p) const { |
| 179 |
*(p) = _data.x; |
| 180 |
*(p + 1) = _data.y; |
| 181 |
} |
| 182 |
|
| 183 |
KOKKOS_INLINE_FUNCTION |
| 184 |
value_type &operator[](const int &i) const { return reinterpret_cast<value_type *>(&_data)[i]; } |
| 185 |
}; |
| 186 |
|
| 187 |
template <> |
| 188 |
class Vector<SIMD<double>, 2> { |
| 189 |
public: |
| 190 |
using type = Vector<SIMD<double>, 2>; |
| 191 |
using value_type = double; |
| 192 |
using mag_type = double; |
| 193 |
|
| 194 |
enum : int { vector_length = 2 }; |
| 195 |
typedef double2 data_type; |
| 196 |
|
| 197 |
KOKKOS_INLINE_FUNCTION |
| 198 |
static const char *label() { return "GpuDouble2"; } |
| 199 |
|
| 200 |
template <typename, int> |
| 201 |
friend class Vector; |
| 202 |
|
| 203 |
private: |
| 204 |
mutable data_type _data; |
| 205 |
|
| 206 |
public: |
| 207 |
KOKKOS_INLINE_FUNCTION Vector() { |
| 208 |
_data.x = 0; |
| 209 |
_data.y = 0; |
| 210 |
} |
| 211 |
KOKKOS_INLINE_FUNCTION Vector(const value_type &val) { |
| 212 |
_data.x = val; |
| 213 |
_data.y = val; |
| 214 |
} |
| 215 |
KOKKOS_INLINE_FUNCTION Vector(const type &b) { |
| 216 |
_data.x = b._data.x; |
| 217 |
_data.y = b._data.y; |
| 218 |
} |
| 219 |
KOKKOS_INLINE_FUNCTION Vector(const double2 &val) { |
| 220 |
_data.x = val.x; |
| 221 |
_data.y = val.y; |
| 222 |
} |
| 223 |
|
| 224 |
template <typename ArgValueType> |
| 225 |
KOKKOS_INLINE_FUNCTION Vector(const ArgValueType &val) { |
| 226 |
_data.x = val; |
| 227 |
_data.y = val; |
| 228 |
} |
| 229 |
|
| 230 |
template <typename ArgValueType> |
| 231 |
KOKKOS_INLINE_FUNCTION Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 232 |
_data.x = b[0]; |
| 233 |
_data.y = b[1]; |
| 234 |
} |
| 235 |
|
| 236 |
KOKKOS_INLINE_FUNCTION |
| 237 |
type &operator=(const double2 &val) { |
| 238 |
_data.x = val.x; |
| 239 |
_data.y = val.y; |
| 240 |
return *this; |
| 241 |
} |
| 242 |
|
| 243 |
KOKKOS_INLINE_FUNCTION |
| 244 |
double2 double2() const { return _data; } |
| 245 |
|
| 246 |
KOKKOS_INLINE_FUNCTION |
| 247 |
type &loadAligned(const value_type *p) { |
| 248 |
_data.x = *(p); |
| 249 |
_data.y = *(p + 1); |
| 250 |
return *this; |
| 251 |
} |
| 252 |
|
| 253 |
KOKKOS_INLINE_FUNCTION |
| 254 |
type &loadUnaligned(const value_type *p) { |
| 255 |
_data.x = *(p); |
| 256 |
_data.y = *(p + 1); |
| 257 |
return *this; |
| 258 |
} |
| 259 |
|
| 260 |
KOKKOS_INLINE_FUNCTION |
| 261 |
void storeAligned(value_type *p) const { |
| 262 |
*(p) = _data.x; |
| 263 |
*(p + 1) = _data.y; |
| 264 |
} |
| 265 |
|
| 266 |
KOKKOS_INLINE_FUNCTION |
| 267 |
void storeUnaligned(value_type *p) const { |
| 268 |
*(p) = _data.x; |
| 269 |
*(p + 1) = _data.y; |
| 270 |
} |
| 271 |
|
| 272 |
KOKKOS_INLINE_FUNCTION |
| 273 |
value_type &operator[](const int &i) const { return reinterpret_cast<value_type *>(&_data)[i]; } |
| 274 |
}; |
| 275 |
|
| 276 |
template <> |
| 277 |
class Vector<SIMD<float>, 4> { |
| 278 |
public: |
| 279 |
using type = Vector<SIMD<float>, 4>; |
| 280 |
using value_type = float; |
| 281 |
using mag_type = float; |
| 282 |
|
| 283 |
enum : int { vector_length = 4 }; |
| 284 |
typedef float4 data_type; |
| 285 |
|
| 286 |
KOKKOS_INLINE_FUNCTION |
| 287 |
static const char *label() { return "GpuFloat4"; } |
| 288 |
|
| 289 |
template <typename, int> |
| 290 |
friend class Vector; |
| 291 |
|
| 292 |
private: |
| 293 |
mutable data_type _data; |
| 294 |
|
| 295 |
public: |
| 296 |
KOKKOS_INLINE_FUNCTION Vector() { |
| 297 |
_data.x = 0; |
| 298 |
_data.y = 0; |
| 299 |
_data.z = 0; |
| 300 |
_data.w = 0; |
| 301 |
} |
| 302 |
KOKKOS_INLINE_FUNCTION Vector(const value_type &val) { |
| 303 |
_data.x = val; |
| 304 |
_data.y = val; |
| 305 |
_data.z = val; |
| 306 |
_data.w = val; |
| 307 |
} |
| 308 |
KOKKOS_INLINE_FUNCTION Vector(const type &b) { |
| 309 |
_data.x = b._data.x; |
| 310 |
_data.y = b._data.y; |
| 311 |
_data.z = b._data.z; |
| 312 |
_data.w = b._data.w; |
| 313 |
} |
| 314 |
KOKKOS_INLINE_FUNCTION Vector(const float4 &val) { |
| 315 |
_data.x = val.x; |
| 316 |
_data.y = val.y; |
| 317 |
_data.z = val.z; |
| 318 |
_data.w = val.w; |
| 319 |
} |
| 320 |
|
| 321 |
template <typename ArgValueType> |
| 322 |
KOKKOS_INLINE_FUNCTION Vector(const ArgValueType &val) { |
| 323 |
_data.x = val; |
| 324 |
_data.y = val; |
| 325 |
_data.z = val; |
| 326 |
_data.w = val; |
| 327 |
} |
| 328 |
|
| 329 |
template <typename ArgValueType> |
| 330 |
KOKKOS_INLINE_FUNCTION Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 331 |
_data.x = b[0]; |
| 332 |
_data.y = b[1]; |
| 333 |
_data.z = b[2]; |
| 334 |
_data.w = b[3]; |
| 335 |
} |
| 336 |
|
| 337 |
KOKKOS_INLINE_FUNCTION |
| 338 |
type &operator=(const float4 &val) { |
| 339 |
_data.x = val.x; |
| 340 |
_data.y = val.y; |
| 341 |
_data.z = val.z; |
| 342 |
_data.w = val.w; |
| 343 |
return *this; |
| 344 |
} |
| 345 |
|
| 346 |
KOKKOS_INLINE_FUNCTION |
| 347 |
float4 float4() const { return _data; } |
| 348 |
|
| 349 |
KOKKOS_INLINE_FUNCTION |
| 350 |
type &loadAligned(const value_type *p) { |
| 351 |
_data.x = *(p); |
| 352 |
_data.y = *(p + 1); |
| 353 |
_data.z = *(p + 2); |
| 354 |
_data.w = *(p + 3); |
| 355 |
return *this; |
| 356 |
} |
| 357 |
|
| 358 |
KOKKOS_INLINE_FUNCTION |
| 359 |
type &loadUnaligned(const value_type *p) { |
| 360 |
_data.x = *(p); |
| 361 |
_data.y = *(p + 1); |
| 362 |
_data.z = *(p + 2); |
| 363 |
_data.w = *(p + 3); |
| 364 |
return *this; |
| 365 |
} |
| 366 |
|
| 367 |
KOKKOS_INLINE_FUNCTION |
| 368 |
void storeAligned(value_type *p) const { |
| 369 |
*(p) = _data.x; |
| 370 |
*(p + 1) = _data.y; |
| 371 |
*(p + 2) = _data.z; |
| 372 |
*(p + 3) = _data.w; |
| 373 |
} |
| 374 |
|
| 375 |
KOKKOS_INLINE_FUNCTION |
| 376 |
void storeUnaligned(value_type *p) const { |
| 377 |
*(p) = _data.x; |
| 378 |
*(p + 1) = _data.y; |
| 379 |
*(p + 2) = _data.z; |
| 380 |
*(p + 3) = _data.w; |
| 381 |
} |
| 382 |
|
| 383 |
KOKKOS_INLINE_FUNCTION |
| 384 |
value_type &operator[](const int &i) const { return reinterpret_cast<value_type *>(&_data)[i]; } |
| 385 |
}; |
| 386 |
|
| 387 |
template <> |
| 388 |
class Vector<SIMD<double>, 4> { |
| 389 |
public: |
| 390 |
using type = Vector<SIMD<double>, 4>; |
| 391 |
using value_type = double; |
| 392 |
using mag_type = double; |
| 393 |
|
| 394 |
enum : int { vector_length = 4 }; |
| 395 |
typedef double4 data_type; |
| 396 |
|
| 397 |
KOKKOS_INLINE_FUNCTION |
| 398 |
static const char *label() { return "GpuDouble4"; } |
| 399 |
|
| 400 |
template <typename, int> |
| 401 |
friend class Vector; |
| 402 |
|
| 403 |
private: |
| 404 |
mutable data_type _data; |
| 405 |
|
| 406 |
public: |
| 407 |
KOKKOS_INLINE_FUNCTION Vector() { |
| 408 |
_data.x = 0; |
| 409 |
_data.y = 0; |
| 410 |
_data.z = 0; |
| 411 |
_data.w = 0; |
| 412 |
} |
| 413 |
KOKKOS_INLINE_FUNCTION Vector(const value_type &val) { |
| 414 |
_data.x = val; |
| 415 |
_data.y = val; |
| 416 |
_data.z = val; |
| 417 |
_data.w = val; |
| 418 |
} |
| 419 |
KOKKOS_INLINE_FUNCTION Vector(const type &b) { |
| 420 |
_data.x = b._data.x; |
| 421 |
_data.y = b._data.y; |
| 422 |
_data.z = b._data.z; |
| 423 |
_data.w = b._data.w; |
| 424 |
} |
| 425 |
KOKKOS_INLINE_FUNCTION Vector(const double4 &val) { |
| 426 |
_data.x = val.x; |
| 427 |
_data.y = val.y; |
| 428 |
_data.z = val.z; |
| 429 |
_data.w = val.w; |
| 430 |
} |
| 431 |
|
| 432 |
template <typename ArgValueType> |
| 433 |
KOKKOS_INLINE_FUNCTION Vector(const ArgValueType &val) { |
| 434 |
_data.x = val; |
| 435 |
_data.y = val; |
| 436 |
_data.z = val; |
| 437 |
_data.w = val; |
| 438 |
} |
| 439 |
|
| 440 |
template <typename ArgValueType> |
| 441 |
KOKKOS_INLINE_FUNCTION Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 442 |
_data.x = b[0]; |
| 443 |
_data.y = b[1]; |
| 444 |
_data.z = b[2]; |
| 445 |
_data.w = b[3]; |
| 446 |
} |
| 447 |
|
| 448 |
KOKKOS_INLINE_FUNCTION |
| 449 |
type &operator=(const double4 &val) { |
| 450 |
_data.x = val.x; |
| 451 |
_data.y = val.y; |
| 452 |
_data.z = val.z; |
| 453 |
_data.w = val.w; |
| 454 |
return *this; |
| 455 |
} |
| 456 |
|
| 457 |
KOKKOS_INLINE_FUNCTION |
| 458 |
double4 double4() const { return _data; } |
| 459 |
|
| 460 |
KOKKOS_INLINE_FUNCTION |
| 461 |
type &loadAligned(const value_type *p) { |
| 462 |
_data.x = *(p); |
| 463 |
_data.y = *(p + 1); |
| 464 |
_data.z = *(p + 2); |
| 465 |
_data.w = *(p + 3); |
| 466 |
return *this; |
| 467 |
} |
| 468 |
|
| 469 |
KOKKOS_INLINE_FUNCTION |
| 470 |
type &loadUnaligned(const value_type *p) { |
| 471 |
_data.x = *(p); |
| 472 |
_data.y = *(p + 1); |
| 473 |
_data.z = *(p + 2); |
| 474 |
_data.w = *(p + 3); |
| 475 |
return *this; |
| 476 |
} |
| 477 |
|
| 478 |
KOKKOS_INLINE_FUNCTION |
| 479 |
void storeAligned(value_type *p) const { |
| 480 |
*(p) = _data.x; |
| 481 |
*(p + 1) = _data.y; |
| 482 |
*(p + 2) = _data.z; |
| 483 |
*(p + 3) = _data.w; |
| 484 |
} |
| 485 |
|
| 486 |
KOKKOS_INLINE_FUNCTION |
| 487 |
void storeUnaligned(value_type *p) const { |
| 488 |
*(p) = _data.x; |
| 489 |
*(p + 1) = _data.y; |
| 490 |
*(p + 2) = _data.z; |
| 491 |
*(p + 3) = _data.w; |
| 492 |
} |
| 493 |
|
| 494 |
KOKKOS_INLINE_FUNCTION |
| 495 |
value_type &operator[](const int &i) const { return reinterpret_cast<value_type *>(&_data)[i]; } |
| 496 |
}; |
| 497 |
|
| 498 |
} // namespace KokkosBatched |
| 499 |
#endif |
| 500 |
|
| 501 |
#if defined(__KOKKOSBATCHED_ENABLE_AVX__) |
| 502 |
#if defined(__AVX__) || defined(__AVX2__) |
| 503 |
|
| 504 |
#if CUDA_VERSION < 12022 |
| 505 |
#undef _Float16 |
| 506 |
#endif |
| 507 |
|
| 508 |
#include <immintrin.h> |
| 509 |
|
| 510 |
namespace KokkosBatched { |
| 511 |
|
| 512 |
template <> |
| 513 |
class Vector<SIMD<double>, 4> { |
| 514 |
public: |
| 515 |
using type = Vector<SIMD<double>, 4>; |
| 516 |
using value_type = double; |
| 517 |
using mag_type = double; |
| 518 |
|
| 519 |
enum : int { vector_length = 4 }; |
| 520 |
typedef __m256d data_type __attribute__((aligned(32))); |
| 521 |
|
| 522 |
inline static const char *label() { return "AVX256"; } |
| 523 |
|
| 524 |
template <typename, int> |
| 525 |
friend class Vector; |
| 526 |
|
| 527 |
private: |
| 528 |
mutable data_type _data; |
| 529 |
|
| 530 |
public: |
| 531 |
inline Vector() { _data = _mm256_setzero_pd(); } |
| 532 |
inline Vector(const value_type &val) { _data = _mm256_set1_pd(val); } |
| 533 |
inline Vector(const type &b) { _data = b._data; } |
| 534 |
inline Vector(const __m256d &val) { _data = val; } |
| 535 |
|
| 536 |
template <typename ArgValueType> |
| 537 |
inline Vector(const ArgValueType &val) { |
| 538 |
auto d = reinterpret_cast<value_type *>(&_data); |
| 539 |
KOKKOSKERNELS_FORCE_SIMD |
| 540 |
for (int i = 0; i < vector_length; ++i) d[i] = val; |
| 541 |
} |
| 542 |
|
| 543 |
template <typename ArgValueType> |
| 544 |
inline Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 545 |
auto dd = reinterpret_cast<value_type *>(&_data); |
| 546 |
auto bb = reinterpret_cast<ArgValueType *>(&b._data); |
| 547 |
KOKKOSKERNELS_FORCE_SIMD |
| 548 |
for (int i = 0; i < vector_length; ++i) dd[i] = bb[i]; |
| 549 |
} |
| 550 |
|
| 551 |
inline type &operator=(const __m256d &val) { |
| 552 |
_data = val; |
| 553 |
return *this; |
| 554 |
} |
| 555 |
|
| 556 |
inline operator __m256d() const { return _data; } |
| 557 |
|
| 558 |
inline type &loadAligned(const value_type *p) { |
| 559 |
_data = _mm256_load_pd(p); |
| 560 |
return *this; |
| 561 |
} |
| 562 |
|
| 563 |
inline type &loadUnaligned(const value_type *p) { |
| 564 |
_data = _mm256_loadu_pd(p); |
| 565 |
return *this; |
| 566 |
} |
| 567 |
|
| 568 |
inline void storeAligned(value_type *p) const { _mm256_store_pd(p, _data); } |
| 569 |
|
| 570 |
inline void storeUnaligned(value_type *p) const { _mm256_storeu_pd(p, _data); } |
| 571 |
|
| 572 |
inline value_type &operator[](const int &i) const { return reinterpret_cast<value_type *>(&_data)[i]; } |
| 573 |
}; |
| 574 |
|
| 575 |
template <> |
| 576 |
class Vector<SIMD<Kokkos::complex<double> >, 2> { |
| 577 |
public: |
| 578 |
using type = Vector<SIMD<Kokkos::complex<double> >, 2>; |
| 579 |
using value_type = Kokkos::complex<double>; |
| 580 |
using mag_type = double; |
| 581 |
|
| 582 |
static const int vector_length = 2; |
| 583 |
typedef __m256d data_type __attribute__((aligned(32))); |
| 584 |
|
| 585 |
inline static const char *label() { return "AVX256"; } |
| 586 |
|
| 587 |
template <typename, int> |
| 588 |
friend class Vector; |
| 589 |
|
| 590 |
private: |
| 591 |
mutable data_type _data; |
| 592 |
|
| 593 |
public: |
| 594 |
inline Vector() { _data = _mm256_setzero_pd(); } |
| 595 |
inline Vector(const value_type &val) { |
| 596 |
_data = _mm256_broadcast_pd((const __m128d *)&val); |
| 597 |
KOKKOSKERNELS_GNU_COMPILER_FENCE |
| 598 |
} |
| 599 |
inline Vector(const mag_type &val) { |
| 600 |
const value_type a(val); |
| 601 |
_data = _mm256_broadcast_pd((__m128d const *)&a); |
| 602 |
KOKKOSKERNELS_GNU_COMPILER_FENCE |
| 603 |
} |
| 604 |
inline Vector(const type &b) { _data = b._data; } |
| 605 |
inline Vector(const __m256d &val) { _data = val; } |
| 606 |
|
| 607 |
// template<typename ArgValueType> |
| 608 |
// inline Vector(const ArgValueType val) { |
| 609 |
// #if defined( KOKKOS_ENABLE_PRAGMA_IVDEP ) |
| 610 |
// #pragma ivdep |
| 611 |
// #endif |
| 612 |
// #if defined( KOKKOS_ENABLE_PRAGMA_VECTOR ) |
| 613 |
// #pragma vector always |
| 614 |
// #endif |
| 615 |
// for (int i=0;i<vector_length;++i) |
| 616 |
// _data.d[i] = value_type(val); |
| 617 |
// } |
| 618 |
template <typename ArgValueType> |
| 619 |
inline Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 620 |
auto dd = reinterpret_cast<value_type *>(&_data); |
| 621 |
auto bb = reinterpret_cast<ArgValueType *>(&b._data); |
| 622 |
KOKKOSKERNELS_FORCE_SIMD |
| 623 |
for (int i = 0; i < vector_length; ++i) dd[i] = bb[i]; |
| 624 |
} |
| 625 |
|
| 626 |
inline type &operator=(const __m256d &val) { |
| 627 |
_data = val; |
| 628 |
return *this; |
| 629 |
} |
| 630 |
|
| 631 |
inline operator __m256d() const { return _data; } |
| 632 |
|
| 633 |
inline type &loadAligned(const value_type *p) { |
| 634 |
_data = _mm256_load_pd((mag_type *)p); |
| 635 |
return *this; |
| 636 |
} |
| 637 |
|
| 638 |
inline type &loadUnaligned(const value_type *p) { |
| 639 |
_data = _mm256_loadu_pd((mag_type *)p); |
| 640 |
return *this; |
| 641 |
} |
| 642 |
|
| 643 |
inline void storeAligned(value_type *p) const { _mm256_store_pd((mag_type *)p, _data); } |
| 644 |
|
| 645 |
inline void storeUnaligned(value_type *p) const { _mm256_storeu_pd((mag_type *)p, _data); } |
| 646 |
|
| 647 |
inline value_type &operator[](const int &i) const { return reinterpret_cast<value_type *>(&_data)[i]; } |
| 648 |
}; |
| 649 |
} // namespace KokkosBatched |
| 650 |
#endif /* #if defined(__AVX__) || defined(__AVX2__) */ |
| 651 |
|
| 652 |
#if defined(__AVX512F__) |
| 653 |
#if CUDA_VERSION < 12022 |
| 654 |
#undef _Float16 |
| 655 |
#endif |
| 656 |
#include <immintrin.h> |
| 657 |
|
| 658 |
namespace KokkosBatched { |
| 659 |
|
| 660 |
template <> |
| 661 |
class Vector<SIMD<double>, 8> { |
| 662 |
public: |
| 663 |
using type = Vector<SIMD<double>, 8>; |
| 664 |
using value_type = double; |
| 665 |
using mag_type = double; |
| 666 |
|
| 667 |
enum : int { vector_length = 8 }; |
| 668 |
typedef __m512d data_type __attribute__((aligned(64))); |
| 669 |
|
| 670 |
inline static const char *label() { return "AVX512"; } |
| 671 |
|
| 672 |
template <typename, int> |
| 673 |
friend class Vector; |
| 674 |
|
| 675 |
private: |
| 676 |
mutable data_type _data; |
| 677 |
|
| 678 |
public: |
| 679 |
inline Vector() { _data = _mm512_setzero_pd(); } |
| 680 |
inline Vector(const value_type &val) { _data = _mm512_set1_pd(val); } |
| 681 |
inline Vector(const type &b) { _data = b._data; } |
| 682 |
inline Vector(const __m512d &val) { _data = val; } |
| 683 |
|
| 684 |
template <typename ArgValueType> |
| 685 |
inline Vector(const ArgValueType &val) { |
| 686 |
auto d = reinterpret_cast<value_type *>(&_data); |
| 687 |
KOKKOSKERNELS_FORCE_SIMD |
| 688 |
for (int i = 0; i < vector_length; ++i) d[i] = val; |
| 689 |
} |
| 690 |
template <typename ArgValueType> |
| 691 |
inline Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 692 |
auto dd = reinterpret_cast<value_type *>(&_data); |
| 693 |
auto bb = reinterpret_cast<ArgValueType *>(&b._data); |
| 694 |
KOKKOSKERNELS_FORCE_SIMD |
| 695 |
for (int i = 0; i < vector_length; ++i) dd[i] = bb[i]; |
| 696 |
} |
| 697 |
|
| 698 |
inline type &operator=(const __m512d &val) { |
| 699 |
_data = val; |
| 700 |
return *this; |
| 701 |
} |
| 702 |
|
| 703 |
inline operator __m512d() const { return _data; } |
| 704 |
|
| 705 |
inline type &loadAligned(const value_type *p) { |
| 706 |
_data = _mm512_load_pd(p); |
| 707 |
return *this; |
| 708 |
} |
| 709 |
|
| 710 |
inline type &loadUnaligned(const value_type *p) { |
| 711 |
_data = _mm512_loadu_pd(p); |
| 712 |
return *this; |
| 713 |
} |
| 714 |
|
| 715 |
inline void storeAligned(value_type *p) const { _mm512_store_pd(p, _data); } |
| 716 |
|
| 717 |
inline void storeUnaligned(value_type *p) const { _mm512_storeu_pd(p, _data); } |
| 718 |
|
| 719 |
inline value_type &operator[](const int &i) const { return reinterpret_cast<value_type *>(&_data)[i]; } |
| 720 |
}; |
| 721 |
|
| 722 |
template <> |
| 723 |
class Vector<SIMD<Kokkos::complex<double> >, 4> { |
| 724 |
public: |
| 725 |
using type = Vector<SIMD<Kokkos::complex<double> >, 4>; |
| 726 |
using value_type = Kokkos::complex<double>; |
| 727 |
using mag_type = double; |
| 728 |
|
| 729 |
enum : int { vector_length = 4 }; |
| 730 |
typedef __m512d data_type __attribute__((aligned(64))); |
| 731 |
|
| 732 |
inline static const char *label() { return "AVX512"; } |
| 733 |
|
| 734 |
template <typename, int> |
| 735 |
friend class Vector; |
| 736 |
|
| 737 |
private: |
| 738 |
mutable data_type _data; |
| 739 |
|
| 740 |
public: |
| 741 |
inline Vector() { _data = _mm512_setzero_pd(); } |
| 742 |
inline Vector(const value_type &val) { |
| 743 |
_data = _mm512_mask_broadcast_f64x4(_mm512_set1_pd(val.imag()), 0x55, _mm256_set1_pd(val.real())); |
| 744 |
KOKKOSKERNELS_GNU_COMPILER_FENCE |
| 745 |
} |
| 746 |
inline Vector(const mag_type &val) { |
| 747 |
_data = _mm512_mask_broadcast_f64x4(_mm512_setzero_pd(), 0x55, _mm256_set1_pd(val)); |
| 748 |
KOKKOSKERNELS_GNU_COMPILER_FENCE |
| 749 |
} |
| 750 |
inline Vector(const type &b) { _data = b._data; } |
| 751 |
inline Vector(const __m512d &val) { _data = val; } |
| 752 |
|
| 753 |
template <typename ArgValueType> |
| 754 |
inline Vector(const ArgValueType &val) { |
| 755 |
auto d = reinterpret_cast<value_type *>(&_data); |
| 756 |
KOKKOSKERNELS_FORCE_SIMD |
| 757 |
for (int i = 0; i < vector_length; ++i) d[i] = val; |
| 758 |
} |
| 759 |
template <typename ArgValueType> |
| 760 |
inline Vector(const Vector<SIMD<ArgValueType>, vector_length> &b) { |
| 761 |
auto dd = reinterpret_cast<value_type *>(&_data); |
| 762 |
auto bb = reinterpret_cast<value_type *>(&b._data); |
| 763 |
KOKKOSKERNELS_FORCE_SIMD |
| 764 |
for (int i = 0; i < vector_length; ++i) dd[i] = bb[i]; |
| 765 |
} |
| 766 |
|
| 767 |
inline type &operator=(const __m512d &val) { |
| 768 |
_data = val; |
| 769 |
return *this; |
| 770 |
} |
| 771 |
|
| 772 |
inline operator __m512d() const { return _data; } |
| 773 |
|
| 774 |
inline type &loadAligned(const value_type *p) { |
| 775 |
_data = _mm512_load_pd((mag_type *)p); |
| 776 |
return *this; |
| 777 |
} |
| 778 |
|
| 779 |
inline type &loadUnaligned(const value_type *p) { |
| 780 |
_data = _mm512_loadu_pd((mag_type *)p); |
| 781 |
return *this; |
| 782 |
} |
| 783 |
|
| 784 |
inline void storeAligned(value_type *p) const { _mm512_store_pd((mag_type *)p, _data); } |
| 785 |
|
| 786 |
inline void storeUnaligned(value_type *p) const { _mm512_storeu_pd((mag_type *)p, _data); } |
| 787 |
|
| 788 |
inline value_type &operator[](const int &i) const { return reinterpret_cast<value_type *>(&_data)[i]; } |
| 789 |
}; |
| 790 |
} // namespace KokkosBatched |
| 791 |
|
| 792 |
#endif /* #if defined(__AVX512F__) */ |
| 793 |
#endif /* #if defined(__KOKKOSBATCHED_ENABLE_AVX__) */ |
| 794 |
|
| 795 |
#include "KokkosBatched_Vector_SIMD_Arith.hpp" |
| 796 |
#include "KokkosBatched_Vector_SIMD_Logical.hpp" |
| 797 |
#include "KokkosBatched_Vector_SIMD_Relation.hpp" |
| 798 |
#include "KokkosBatched_Vector_SIMD_Math.hpp" |
| 799 |
#include "KokkosBatched_Vector_SIMD_Misc.hpp" |
| 800 |
#include "KokkosBatched_Vector_SIMD_View.hpp" |
| 801 |
|
| 802 |
#endif |