float16.h 19.0 KB
Newer Older
K
Kexin Zhao 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

K
Kexin Zhao 已提交
15 16
// need to define PADDLE_ARM_FP16

K
Kexin Zhao 已提交
17 18 19 20 21 22
#pragma once

#include <cstdint>
#include <istream>
#include <ostream>

K
Kexin Zhao 已提交
23
#define USE_EIGEN
K
Kexin Zhao 已提交
24 25

#ifdef USE_EIGEN  // delete this #if macro
K
Kexin Zhao 已提交
26
#include "unsupported/Eigen/CXX11/Tensor"
K
Kexin Zhao 已提交
27 28
#endif

K
Kexin Zhao 已提交
29 30 31 32 33 34 35 36 37 38 39 40
#ifdef __GNUC__
#define PADDLE_GNUC_VER (__GNUC__ * 10 + __GNUC_MINOR__)
#else
#define PADDLE_GNUC_VER 0
#endif  // __GNUC__

#ifdef __clang__
#define PADDLE_CLANG_VER (__clang_major__ * 10 + __clang_minor__)
#else
#define PADDLE_CLANG_VER 0
#endif  // __clang__

K
Kexin Zhao 已提交
41 42 43 44 45 46 47 48
#ifdef __CUDACC__
#define PADDLE_HOSTDEVICE __host__ __device__
#if CUDA_VERSION >= 7050
#define PADDLE_CUDA_FP16
#include <cuda_fp16.h>
#endif  // CUDA_VERSION >= 7050
#else
#define PADDLE_HOSTDEVICE
K
Kexin Zhao 已提交
49
#endif  // __CUDACC__
K
Kexin Zhao 已提交
50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

#ifdef __arm__
#define PADDLE_ARM_32
#endif

#ifdef __aarch64__
#define PADDLE_ARM_64
#endif

#if defined(PADDLE_ARM_32) || defined(PADDLE_ARM_64)
#define PADDLE_ARM
#endif

#if defined(__ARM_NEON) || defined(__ARM_NEON__)
#define PADDLE_NEON
K
Kexin Zhao 已提交
65
#include <arm_neon.h>
K
Kexin Zhao 已提交
66 67 68 69 70 71 72 73 74 75
#endif

#if defined(PADDLE_NEON) && defined(PADDLE_ARM_32)
#define PADDLE_NEON_32
#endif

#if defined(PADDLE_NEON) && defined(PADDLE_ARM_64)
#define PADDLE_NEON_64
#endif

K
Kexin Zhao 已提交
76
#ifdef PADDLE_ARM
K
Kexin Zhao 已提交
77 78
#ifdef __F16C__
#undef __F16C__
K
Kexin Zhao 已提交
79 80 81 82
#endif  // __F16C__
#else
#include <immintrin.h>
#endif  // PADDLE_ARM
K
Kexin Zhao 已提交
83

K
Kexin Zhao 已提交
84
#define PADDLE_ALIGN(x) __attribute__((aligned(x)))
K
Kexin Zhao 已提交
85 86 87

namespace paddle {

K
Kexin Zhao 已提交
88
struct float16;
K
Kexin Zhao 已提交
89

K
Kexin Zhao 已提交
90
namespace fp16_impl {
K
Kexin Zhao 已提交
91
// convert from float to half precision in round-to-nearest-even mode
K
Kexin Zhao 已提交
92 93 94
PADDLE_HOSTDEVICE inline float16 float_to_half_rn(float f);
PADDLE_HOSTDEVICE inline float half_to_float(float16 h);
}  // namespace fp16_impl
K
Kexin Zhao 已提交
95

K
Kexin Zhao 已提交
96 97 98
// Use PADDLE_ALIGNED(2) to ensure that each float16 will be allocated
// and aligned at least on a 2-byte boundary, which leads to efficient
// memory access of float16 struct and also makes float16 compatible
K
Kexin Zhao 已提交
99
// with CUDA half, ARM float16_t, and Eigen::half data types.
K
Kexin Zhao 已提交
100 101
struct PADDLE_ALIGN(2) float16 {
  uint16_t x;
K
Kexin Zhao 已提交
102

K
Kexin Zhao 已提交
103 104 105 106 107
  PADDLE_HOSTDEVICE inline float16() {}

  PADDLE_HOSTDEVICE inline float16(const float16& h) : x(h.x) {}

#ifdef PADDLE_CUDA_FP16
K
Kexin Zhao 已提交
108
  PADDLE_HOSTDEVICE inline float16(const half& h) {
K
Kexin Zhao 已提交
109 110 111 112 113 114 115 116 117 118 119 120
#if CUDA_VERSION >= 9000
    x = reinterpret_cast<__half_raw*>(&h)->x;
#else
    x = h.x;
#endif  // CUDA_VERSION >= 9000
  }
#endif  // PADDLE_CUDA_FP16

#ifdef USE_EIGEN
  PADDLE_HOSTDEVICE inline float16(const Eigen::half& h) : x(h.x) {}
#endif  // USE_EIGEN

K
Kexin Zhao 已提交
121 122
#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \
    (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34)
K
Kexin Zhao 已提交
123
  // __fp16 is a native half precision data type for arm cpu,
K
Kexin Zhao 已提交
124 125 126 127 128
  // float16_t is an alias for __fp16 in arm_fp16.h,
  // which is included in arm_neon.h.
  // According to gcc, __fp16 can only be used as an argument to fp16
  // intrinsic defined in arm_neon.h or as a storage type. It cannot
  // be used as a formal function argument.
K
Kexin Zhao 已提交
129
  // TODO(kexinzhao): test it on RPI
K
Kexin Zhao 已提交
130 131
  PADDLE_HOSTDEVICE inline float16(const float16_t* h) {
    x = *reinterpret_cast<uint16_t*>(h);
K
Kexin Zhao 已提交
132 133 134 135
  }
#endif

  PADDLE_HOSTDEVICE inline explicit float16(bool b) : x(b ? 0x3c00 : 0) {}
K
Kexin Zhao 已提交
136

K
Kexin Zhao 已提交
137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176
  PADDLE_HOSTDEVICE inline explicit float16(int8_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
  }

  PADDLE_HOSTDEVICE inline explicit float16(uint8_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
  }

  PADDLE_HOSTDEVICE inline explicit float16(int16_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
  }

  PADDLE_HOSTDEVICE inline explicit float16(uint16_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
  }

  PADDLE_HOSTDEVICE inline explicit float16(int32_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
  }

  PADDLE_HOSTDEVICE inline explicit float16(uint32_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
  }

  PADDLE_HOSTDEVICE inline explicit float16(int64_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
  }

  PADDLE_HOSTDEVICE inline explicit float16(uint64_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
  }

K
Kexin Zhao 已提交
177
  PADDLE_HOSTDEVICE inline explicit float16(float val) {
K
Kexin Zhao 已提交
178 179 180 181
    float16 res = fp16_impl::float_to_half_rn(val);
    x = res.x;
  }

K
Kexin Zhao 已提交
182
  PADDLE_HOSTDEVICE inline explicit float16(double val) {
K
Kexin Zhao 已提交
183
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
K
Kexin Zhao 已提交
184 185 186
    x = res.x;
  }

K
Kexin Zhao 已提交
187 188 189 190 191 192
  PADDLE_HOSTDEVICE inline float16& operator=(const float16& rhs) {
    x = rhs.x;
    return *this;
  }

#ifdef PADDLE_CUDA_FP16
K
Kexin Zhao 已提交
193
  PADDLE_HOSTDEVICE inline float16& operator=(const half& rhs) {
K
Kexin Zhao 已提交
194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209
#if CUDA_VERSION >= 9000
    x = reinterpret_cast<__half_raw*>(&rhs)->x;
#else
    x = rhs.x;
#endif
    return *this;
  }
#endif

#ifdef USE_EIGEN
  PADDLE_HOSTDEVICE inline float16& operator=(const Eigen::half& rhs) {
    x = rhs.x;
    return *this;
  }
#endif  // USE_EIGEN

K
Kexin Zhao 已提交
210 211
#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \
    (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34)
K
Kexin Zhao 已提交
212 213
  PADDLE_HOSTDEVICE inline float16& operator=(const float16_t* rhs) {
    x = *reinterpret_cast<uint16_t*>(rhs);
K
Kexin Zhao 已提交
214 215 216 217
    return *this;
  }
#endif

K
Kexin Zhao 已提交
218 219 220 221 222 223
  PADDLE_HOSTDEVICE inline float16& operator=(bool b) {
    x = b ? 0x3c00 : 0;
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(int8_t val) {
K
Kexin Zhao 已提交
224
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
K
Kexin Zhao 已提交
225
    x = res.x;
K
Kexin Zhao 已提交
226
    return *this;
K
Kexin Zhao 已提交
227 228
  }

K
Kexin Zhao 已提交
229
  PADDLE_HOSTDEVICE inline float16& operator=(uint8_t val) {
K
Kexin Zhao 已提交
230
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
K
Kexin Zhao 已提交
231
    x = res.x;
K
Kexin Zhao 已提交
232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(int16_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(uint16_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(int32_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(uint32_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(int64_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(uint64_t val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(float val) {
    float16 res = fp16_impl::float_to_half_rn(val);
    x = res.x;
    return *this;
  }

  PADDLE_HOSTDEVICE inline float16& operator=(double val) {
    float16 res = fp16_impl::float_to_half_rn(static_cast<float>(val));
    x = res.x;
    return *this;
K
Kexin Zhao 已提交
281
  }
K
Kexin Zhao 已提交
282 283

#ifdef PADDLE_CUDA_FP16
K
Kexin Zhao 已提交
284
  PADDLE_HOSTDEVICE inline operator half() const {
K
Kexin Zhao 已提交
285 286 287 288 289 290 291 292 293 294 295
#if CUDA_VERSION >= 9000
    __half_raw h;
    h.x = x;
    return half(h);
#else
    half h;
    h.x = x;
    return h;
#endif  // CUDA_VERSION >= 9000
  }
#endif  // PADDLE_CUDA_FP16
K
Kexin Zhao 已提交
296

K
Kexin Zhao 已提交
297
#ifdef USE_EIGEN
K
Kexin Zhao 已提交
298
  PADDLE_HOSTDEVICE inline operator Eigen::half() const {
K
Kexin Zhao 已提交
299 300 301 302 303 304
    Eigen::half h;
    h.x = x;
    return h;
  }
#endif  // USE_EIGEN

K
Kexin Zhao 已提交
305 306
#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \
    (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34)
K
Kexin Zhao 已提交
307 308
  // check whether it works or not
  PADDLE_HOSTDEVICE inline operator float16_t() const {
K
Kexin Zhao 已提交
309 310 311 312 313
    float16 h = *this;
    return *reinterpret_cast<float16_t*>(&h);
  }
#endif

K
Kexin Zhao 已提交
314
  PADDLE_HOSTDEVICE inline explicit operator bool() const {
K
Kexin Zhao 已提交
315 316 317
    return (x & 0x7fff) != 0;
  }

K
Kexin Zhao 已提交
318 319
  PADDLE_HOSTDEVICE inline explicit operator int8_t() const {
    return static_cast<int8_t>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
320 321
  }

K
Kexin Zhao 已提交
322 323
  PADDLE_HOSTDEVICE inline explicit operator uint8_t() const {
    return static_cast<uint8_t>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
324 325
  }

K
Kexin Zhao 已提交
326 327
  PADDLE_HOSTDEVICE inline explicit operator int16_t() const {
    return static_cast<int16_t>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
328 329
  }

K
Kexin Zhao 已提交
330 331
  PADDLE_HOSTDEVICE inline explicit operator uint16_t() const {
    return static_cast<uint16_t>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
332 333
  }

K
Kexin Zhao 已提交
334 335
  PADDLE_HOSTDEVICE inline explicit operator int32_t() const {
    return static_cast<int32_t>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
336 337
  }

K
Kexin Zhao 已提交
338 339
  PADDLE_HOSTDEVICE inline explicit operator uint32_t() const {
    return static_cast<uint32_t>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
340 341
  }

K
Kexin Zhao 已提交
342 343
  PADDLE_HOSTDEVICE inline explicit operator int64_t() const {
    return static_cast<int64_t>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
344 345
  }

K
Kexin Zhao 已提交
346 347
  PADDLE_HOSTDEVICE inline explicit operator uint64_t() const {
    return static_cast<uint64_t>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
348 349
  }

K
Kexin Zhao 已提交
350
  PADDLE_HOSTDEVICE inline explicit operator float() const {
K
Kexin Zhao 已提交
351 352 353
    return fp16_impl::half_to_float(*this);
  }

K
Kexin Zhao 已提交
354 355
  PADDLE_HOSTDEVICE inline explicit operator double() const {
    return static_cast<double>(fp16_impl::half_to_float(*this));
K
Kexin Zhao 已提交
356 357 358 359 360 361
  }
};

// arithmetic operators
#if defined(PADDLE_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
__device__ inline float16 operator+(const float16& a, const float16& b) {
K
Kexin Zhao 已提交
362
  return float16(__hadd(half(a), half(b)));
K
Kexin Zhao 已提交
363 364 365
}

__device__ inline float16 operator-(const float16& a, const float16& b) {
K
Kexin Zhao 已提交
366
  return float16(__hsub(half(a), half(b)));
K
Kexin Zhao 已提交
367 368 369
}

__device__ inline float16 operator*(const float16& a, const float16& b) {
K
Kexin Zhao 已提交
370
  return float16(__hmul(half(a), half(b)));
K
Kexin Zhao 已提交
371 372
}

K
Kexin Zhao 已提交
373 374 375 376 377 378
__device__ inline float16 operator/(const float16& a, const float16& b) {
  // TODO(kexinzhao): check the cuda version that starts to support __hdiv
  float num = __half2float(half(a));
  float denom = __half2float(half(b));
  return float16(num / denom);
}
K
Kexin Zhao 已提交
379

K
Kexin Zhao 已提交
380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428
__device__ inline float16 operator-(const float16& a) {
  return float16(__hneg(half(a)));
}

__device__ inline float16& operator+=(float16& a, const float16& b) {
  a = a + b;
  return a;
}

__device__ inline float16& operator-=(float16& a, const float16& b) {
  a = a - b;
  return a;
}

__device__ inline float16& operator*=(float16& a, const float16& b) {
  a = a * b;
  return a;
}

__device__ inline float16& operator/=(float16& a, const float16& b) {
  a = a / b;
  return a;
}

__device__ inline bool operator==(const float16& a, const float16& b) {
  return __heq(half(a), half(b));
}

__device__ inline bool operator!=(const float16& a, const float16& b) {
  return __hne(half(a), half(b));
}

__device__ inline bool operator<(const float16& a, const float16& b) {
  return __hlt(half(a), half(b));
}

__device__ inline bool operator<=(const float16& a, const float16& b) {
  return __hle(half(a), half(b));
}

__device__ inline bool operator>(const float16& a, const float16& b) {
  return __hgt(half(a), half(b));
}

__device__ inline bool operator>=(const float16& a, const float16& b) {
  return __hge(half(a), half(b));
}

// On ARMv8.2-A CPU
K
Kexin Zhao 已提交
429 430
#elif defined(PADDLE_NEON_64) && defined(PADDLE_ARM_FP16) && \
    (PADDLE_GNUC_VER >= 71 || PADDLE_CLANG_VER >= 39)
K
Kexin Zhao 已提交
431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561
__host__ inline float16 operator+(const float16& a, const float16& b) {
  return float16(vaddh_f16(float16_t(a), float16_t(b)));
}

__host__ inline float16 operator-(const float16& a, const float16& b) {
  return float16(vsubh_f16(float16_t(a), float16_t(b)));
}

__host__ inline float16 operator*(const float16& a, const float16& b) {
  return float16(vmulh_f16(float16_t(a), float16_t(b)));
}

__host__ inline float16 operator/(const float16& a, const float16& b) {
  return float16(vdivh_f16(float16_t(a), float16_t(b)));
}

__host__ inline float16 operator-(const float16& a) {
  return float16(vnegh_f16(float16_t(a)));
}

__host__ inline float16& operator+=(float16& a, const float16& b) {
  a = a + b;
  return a;
}

__host__ inline float16& operator-=(float16& a, const float16& b) {
  a = a - b;
  return a;
}

__host__ inline float16& operator*=(float16& a, const float16& b) {
  a = a * b;
  return a;
}

__host__ inline float16& operator/=(float16& a, const float16& b) {
  a = a / b;
  return a;
}

__host__ inline bool operator==(const float16& a, const float16& b) {
  return static_cast<bool>(vceqh_f16(float16_t(a), float16_t(b)));
}

__host__ inline bool operator!=(const float16& a, const float16& b) {
  return !(a == b);
}

// compare only available in NEON_64
__host__ inline bool operator<(const float16& a, const float16& b) {
  return static_cast<bool>(vclth_f16(float16_t(a), float16_t(b)));
}

__host__ inline bool operator<=(const float16& a, const float16& b) {
  return static_cast<bool>(vcleh_f16(float16_t(a), float16_t(b)));
}

__host__ inline bool operator>(const float16& a, const float16& b) {
  return static_cast<bool>(vcgth_f16(float16_t(a), float16_t(b)));
}

__host__ inline bool operator>=(const float16& a, const float16& b) {
  return static_cast<bool>(vcgeh_f16(float16_t(a), float16_t(b)));
}

#else  // software emulation on other cpu
PADDLE_HOSTDEVICE inline float16 operator+(const float16& a, const float16& b) {
  return float16(float(a) + float(b));
}

PADDLE_HOSTDEVICE inline float16 operator-(const float16& a, const float16& b) {
  return float16(float(a) - float(b));
}

PADDLE_HOSTDEVICE inline float16 operator*(const float16& a, const float16& b) {
  return float16(float(a) * float(b));
}

PADDLE_HOSTDEVICE inline float16 operator/(const float16& a, const float16& b) {
  return float16(float(a) / float(b));
}

PADDLE_HOSTDEVICE inline float16 operator-(const float16& a) {
  float16 res;
  res.x = a.x ^ 0x8000;
  return res;
}

PADDLE_HOSTDEVICE inline float16& operator+=(float16& a, const float16& b) {
  a = float16(float(a) + float(b));
  return a;
}

PADDLE_HOSTDEVICE inline float16& operator-=(float16& a, const float16& b) {
  a = float16(float(a) - float(b));
  return a;
}

PADDLE_HOSTDEVICE inline float16& operator*=(float16& a, const float16& b) {
  a = float16(float(a) * float(b));
  return a;
}

PADDLE_HOSTDEVICE inline float16& operator/=(float16& a, const float16& b) {
  a = float16(float(a) / float(b));
  return a;
}

PADDLE_HOSTDEVICE inline bool operator==(const float16& a, const float16& b) {
  return float(a) == float(b);
}

PADDLE_HOSTDEVICE inline bool operator!=(const float16& a, const float16& b) {
  return float(a) != float(b);
}

PADDLE_HOSTDEVICE inline bool operator<(const float16& a, const float16& b) {
  return float(a) < float(b);
}

PADDLE_HOSTDEVICE inline bool operator<=(const float16& a, const float16& b) {
  return float(a) <= float(b);
}

PADDLE_HOSTDEVICE inline bool operator>(const float16& a, const float16& b) {
  return float(a) > float(b);
}

PADDLE_HOSTDEVICE inline bool operator>=(const float16& a, const float16& b) {
  return float(a) >= float(b);
}
K
Kexin Zhao 已提交
562 563

#endif
K
Kexin Zhao 已提交
564 565 566

namespace fp16_impl {

K
Kexin Zhao 已提交
567
union Bits {
K
Kexin Zhao 已提交
568 569 570 571 572
  float f;
  int32_t si;
  uint32_t ui;
};

K
Kexin Zhao 已提交
573 574 575 576 577 578 579 580 581 582 583 584 585 586
const int shift = 13;
const int shiftSign = 16;

const int32_t infN = 0x7F800000;
const int32_t maxN = 0x477FE000;  // max flt16 as flt32
const int32_t minN = 0x38800000;  // min flt16 normal as flt32
const int32_t sigN = 0x80000000;  // sign bit

constexpr int32_t infC = infN >> shift;
constexpr int32_t nanN = (infC + 1) << shift;  // minimum flt16 nan as float32
constexpr int32_t maxC = maxN >> shift;
constexpr int32_t minC = minN >> shift;
constexpr int32_t sigC = sigN >> shiftSign;

K
Kexin Zhao 已提交
587
const int32_t mulN = 0x52000000;  // (1 << 23) / minN
K
Kexin Zhao 已提交
588 589 590 591 592 593 594 595 596 597
const int32_t mulC = 0x33800000;  // minN / (1 << (23 - shift))
const int32_t subC = 0x003FF;     // max flt32 subnormal downshifted
const int32_t norC = 0x00400;     // min flt32 normal downshifted

constexpr int32_t maxD = infC - maxC - 1;
constexpr int32_t minD = minC - subC - 1;

PADDLE_HOSTDEVICE inline float16 float_to_half_rn(float f) {
#if defined(PADDLE_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
  half tmp = __float2half(f);
K
Kexin Zhao 已提交
598
  return *reinterpret_cast<float16*>(&tmp);
K
Kexin Zhao 已提交
599

K
Kexin Zhao 已提交
600
#elif defined(PADDLE_NEON_64)  // test on RPI
K
Kexin Zhao 已提交
601 602 603
  float16 res;
  asm volatile(
      "ld1 {v0.s}[0], [%[float_ptr]]\n"
K
Kexin Zhao 已提交
604
      "fcvt h0, s0\n"
K
Kexin Zhao 已提交
605 606 607 608 609 610 611 612
      "st1 {v0.h}[0], [%[half_ptr]]\n"
      :  // outputs
      :  // inputs
      [float_ptr] "r"(&f),
      [half_ptr] "r"(&(res.x))
      :  // clobbers
      "memory", "v0");
  return res;
K
Kexin Zhao 已提交
613

K
Kexin Zhao 已提交
614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632
#elif defined(PADDLE_NEON_32)  // test on RPI
  float16 res;
  asm volatile(
      "vld1.32 {d0[0]}, [%[float_ptr]]\n"
      "vcvt.f16.f32 d0, q0\n"
      "vst1.16 {d0[0]}, [%[half_ptr]]\n"
      :  // outputs
      :  // inputs
      [float_ptr] "r"(&f),
      [half_ptr] "r"(&(res.x))
      :  // clobbers
      "memory", "d0");
  return res;

#elif defined(__F16C__)
  float16 res;
  res.x = _cvtss_sh(f, 0);
  return res;

K
Kexin Zhao 已提交
633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651
#else
  // Conversion routine adapted from
  // http://stackoverflow.com/questions/1659440/32-bit-to-16-bit-floating-point-conversion
  Bits v, s;
  v.f = f;
  uint32_t sign = v.si & sigN;
  v.si ^= sign;
  sign >>= shiftSign;  // logical shift
  s.si = mulN;
  s.si = s.f * v.f;  // correct subnormals
  v.si ^= (s.si ^ v.si) & -(minN > v.si);
  v.si ^= (infN ^ v.si) & -((infN > v.si) & (v.si > maxN));
  v.si ^= (nanN ^ v.si) & -((nanN > v.si) & (v.si > infN));
  v.ui >>= shift;  // logical shift
  v.si ^= ((v.si - maxD) ^ v.si) & -(v.si > maxC);
  v.si ^= ((v.si - minD) ^ v.si) & -(v.si > subC);
  float16 res;
  res.x = v.ui | sign;
  return res;
K
Kexin Zhao 已提交
652

K
Kexin Zhao 已提交
653 654
#endif
}
K
Kexin Zhao 已提交
655

K
Kexin Zhao 已提交
656 657 658 659 660
PADDLE_HOSTDEVICE inline float half_to_float(float16 h) {
#if defined(PADDLE_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
  half tmp = *reinterpret_cast<half*>(&h);
  return __half2float(h);

K
Kexin Zhao 已提交
661
#elif defined(PADDLE_NEON_64)
K
Kexin Zhao 已提交
662 663 664 665 666 667 668 669 670 671 672 673
  float res;
  asm volatile(
      "ld1 {v0.h}[0], [%[half_ptr]]\n"
      "FCVT s0, h0\n"
      "st1 {v0.s}[0], [%[float_ptr]]\n"
      :  // outputs
      :  // inputs
      [half_ptr] "r"(&(h.x)),
      [float_ptr] "r"(&res)
      :  // clobbers
      "memory", "v0");
  return res;
K
Kexin Zhao 已提交
674

K
Kexin Zhao 已提交
675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691
#elif defined(PADDLE_NEON_32)
  float res;
  asm volatile(
      "vld1.16 {d0[0]}, [%[half_ptr]]\n"
      "vcvt.f32.f16 q0, d0\n"
      "vst1.32 {d0[0]}, [%[float_ptr]]\n"
      :  // outputs
      :  // inputs
      [half_ptr] "r"(&(h.x)),
      [float_ptr] "r"(&res)
      :  // clobbers
      "memory", "v0");
  return res;

#elif defined(__F16C__)
  return _cvtsh_ss(h.x);

K
Kexin Zhao 已提交
692 693 694 695
#else
  // Conversion routine adapted from
  // http://stackoverflow.com/questions/1659440/32-bit-to-16-bit-floating-point-conversion
  Bits v;
K
Kexin Zhao 已提交
696
  v.ui = h.x;
K
Kexin Zhao 已提交
697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713
  int32_t sign = v.si & sigC;
  v.si ^= sign;
  sign <<= shiftSign;
  v.si ^= ((v.si + minD) ^ v.si) & -(v.si > subC);
  v.si ^= ((v.si + maxD) ^ v.si) & -(v.si > maxC);
  Bits s;
  s.si = mulC;
  s.f *= v.si;
  int32_t mask = -(norC > v.si);
  v.si <<= shift;
  v.si ^= (s.si ^ v.si) & mask;
  v.si |= sign;
  return v.f;

#endif
}

K
Kexin Zhao 已提交
714
}  // namespace fp16_impl
K
Kexin Zhao 已提交
715 716

}  // namespace paddle