Home | History | Annotate | Download | only in CUDA
      1 // This file is part of Eigen, a lightweight C++ template library
      2 // for linear algebra.
      3 //
      4 // This Source Code Form is subject to the terms of the Mozilla
      5 // Public License v. 2.0. If a copy of the MPL was not distributed
      6 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
      7 //
      8 // The conversion routines are Copyright (c) Fabian Giesen, 2016.
      9 // The original license follows:
     10 //
     11 // Copyright (c) Fabian Giesen, 2016
     12 // All rights reserved.
     13 // Redistribution and use in source and binary forms, with or without
     14 // modification, are permitted.
     15 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
     16 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
     17 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
     18 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
     19 // HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
     20 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
     21 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
     22 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
     23 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
     24 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
     25 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     26 
     27 
     28 // Standard 16-bit float type, mostly useful for GPUs. Defines a new
     29 // type Eigen::half (inheriting from CUDA's __half struct) with
     30 // operator overloads such that it behaves basically as an arithmetic
     31 // type. It will be quite slow on CPUs (so it is recommended to stay
     32 // in fp32 for CPUs, except for simple parameter conversions, I/O
     33 // to disk and the likes), but fast on GPUs.
     34 
     35 
     36 #ifndef EIGEN_HALF_CUDA_H
     37 #define EIGEN_HALF_CUDA_H
     38 
     39 #if __cplusplus > 199711L
     40 #define EIGEN_EXPLICIT_CAST(tgt_type) explicit operator tgt_type()
     41 #else
     42 #define EIGEN_EXPLICIT_CAST(tgt_type) operator tgt_type()
     43 #endif
     44 
     45 
     46 namespace Eigen {
     47 
     48 struct half;
     49 
     50 namespace half_impl {
     51 
     52 #if !defined(EIGEN_HAS_CUDA_FP16)
     53 
     54 // Make our own __half definition that is similar to CUDA's.
     55 struct __half {
     56   EIGEN_DEVICE_FUNC __half() {}
     57   explicit EIGEN_DEVICE_FUNC __half(unsigned short raw) : x(raw) {}
     58   unsigned short x;
     59 };
     60 
     61 #endif
     62 
     63 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x);
     64 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff);
     65 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h);
     66 
     67 struct half_base : public __half {
     68   EIGEN_DEVICE_FUNC half_base() {}
     69   EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half(h) {}
     70   EIGEN_DEVICE_FUNC half_base(const __half& h) : __half(h) {}
     71 };
     72 
     73 } // namespace half_impl
     74 
     75 // Class definition.
     76 struct half : public half_impl::half_base {
     77   #if !defined(EIGEN_HAS_CUDA_FP16)
     78     typedef half_impl::__half __half;
     79   #endif
     80 
     81   EIGEN_DEVICE_FUNC half() {}
     82 
     83   EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
     84   EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
     85 
     86   explicit EIGEN_DEVICE_FUNC half(bool b)
     87       : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
     88   template<class T>
     89   explicit EIGEN_DEVICE_FUNC half(const T& val)
     90       : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(val))) {}
     91   explicit EIGEN_DEVICE_FUNC half(float f)
     92       : half_impl::half_base(half_impl::float_to_half_rtne(f)) {}
     93 
     94   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(bool) const {
     95     // +0.0 and -0.0 become false, everything else becomes true.
     96     return (x & 0x7fff) != 0;
     97   }
     98   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(signed char) const {
     99     return static_cast<signed char>(half_impl::half_to_float(*this));
    100   }
    101   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned char) const {
    102     return static_cast<unsigned char>(half_impl::half_to_float(*this));
    103   }
    104   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(short) const {
    105     return static_cast<short>(half_impl::half_to_float(*this));
    106   }
    107   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned short) const {
    108     return static_cast<unsigned short>(half_impl::half_to_float(*this));
    109   }
    110   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(int) const {
    111     return static_cast<int>(half_impl::half_to_float(*this));
    112   }
    113   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned int) const {
    114     return static_cast<unsigned int>(half_impl::half_to_float(*this));
    115   }
    116   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long) const {
    117     return static_cast<long>(half_impl::half_to_float(*this));
    118   }
    119   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long) const {
    120     return static_cast<unsigned long>(half_impl::half_to_float(*this));
    121   }
    122   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long long) const {
    123     return static_cast<long long>(half_impl::half_to_float(*this));
    124   }
    125   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long long) const {
    126     return static_cast<unsigned long long>(half_to_float(*this));
    127   }
    128   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(float) const {
    129     return half_impl::half_to_float(*this);
    130   }
    131   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(double) const {
    132     return static_cast<double>(half_impl::half_to_float(*this));
    133   }
    134 
    135   EIGEN_DEVICE_FUNC half& operator=(const half& other) {
    136     x = other.x;
    137     return *this;
    138   }
    139 };
    140 
    141 namespace half_impl {
    142 
    143 #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
    144 
    145 // Intrinsics for native fp16 support. Note that on current hardware,
    146 // these are no faster than fp32 arithmetic (you need to use the half2
    147 // versions to get the ALU speed increased), but you do save the
    148 // conversion steps back and forth.
    149 
    150 __device__ half operator + (const half& a, const half& b) {
    151   return __hadd(a, b);
    152 }
    153 __device__ half operator * (const half& a, const half& b) {
    154   return __hmul(a, b);
    155 }
    156 __device__ half operator - (const half& a, const half& b) {
    157   return __hsub(a, b);
    158 }
    159 __device__ half operator / (const half& a, const half& b) {
    160   float num = __half2float(a);
    161   float denom = __half2float(b);
    162   return __float2half(num / denom);
    163 }
    164 __device__ half operator - (const half& a) {
    165   return __hneg(a);
    166 }
    167 __device__ half& operator += (half& a, const half& b) {
    168   a = a + b;
    169   return a;
    170 }
    171 __device__ half& operator *= (half& a, const half& b) {
    172   a = a * b;
    173   return a;
    174 }
    175 __device__ half& operator -= (half& a, const half& b) {
    176   a = a - b;
    177   return a;
    178 }
    179 __device__ half& operator /= (half& a, const half& b) {
    180   a = a / b;
    181   return a;
    182 }
    183 __device__ bool operator == (const half& a, const half& b) {
    184   return __heq(a, b);
    185 }
    186 __device__ bool operator != (const half& a, const half& b) {
    187   return __hne(a, b);
    188 }
    189 __device__ bool operator < (const half& a, const half& b) {
    190   return __hlt(a, b);
    191 }
    192 __device__ bool operator <= (const half& a, const half& b) {
    193   return __hle(a, b);
    194 }
    195 __device__ bool operator > (const half& a, const half& b) {
    196   return __hgt(a, b);
    197 }
    198 __device__ bool operator >= (const half& a, const half& b) {
    199   return __hge(a, b);
    200 }
    201 
    202 #else  // Emulate support for half floats
    203 
    204 // Definitions for CPUs and older CUDA, mostly working through conversion
    205 // to/from fp32.
    206 
    207 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
    208   return half(float(a) + float(b));
    209 }
    210 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
    211   return half(float(a) * float(b));
    212 }
    213 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
    214   return half(float(a) - float(b));
    215 }
    216 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
    217   return half(float(a) / float(b));
    218 }
    219 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
    220   half result;
    221   result.x = a.x ^ 0x8000;
    222   return result;
    223 }
    224 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
    225   a = half(float(a) + float(b));
    226   return a;
    227 }
    228 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
    229   a = half(float(a) * float(b));
    230   return a;
    231 }
    232 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
    233   a = half(float(a) - float(b));
    234   return a;
    235 }
    236 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
    237   a = half(float(a) / float(b));
    238   return a;
    239 }
    240 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
    241   return float(a) == float(b);
    242 }
    243 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
    244   return float(a) != float(b);
    245 }
    246 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
    247   return float(a) < float(b);
    248 }
    249 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
    250   return float(a) <= float(b);
    251 }
    252 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
    253   return float(a) > float(b);
    254 }
    255 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
    256   return float(a) >= float(b);
    257 }
    258 
    259 #endif  // Emulate support for half floats
    260 
    261 // Division by an index. Do it in full float precision to avoid accuracy
    262 // issues in converting the denominator to half.
    263 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
    264   return half(static_cast<float>(a) / static_cast<float>(b));
    265 }
    266 
    267 // Conversion routines, including fallbacks for the host or older CUDA.
    268 // Note that newer Intel CPUs (Haswell or newer) have vectorized versions of
    269 // these in hardware. If we need more performance on older/other CPUs, they are
    270 // also possible to vectorize directly.
    271 
    272 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x) {
    273   __half h;
    274   h.x = x;
    275   return h;
    276 }
    277 
    278 union FP32 {
    279   unsigned int u;
    280   float f;
    281 };
    282 
    283 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
    284 #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
    285   return __float2half(ff);
    286 
    287 #elif defined(EIGEN_HAS_FP16_C)
    288   __half h;
    289   h.x = _cvtss_sh(ff, 0);
    290   return h;
    291 
    292 #else
    293   FP32 f; f.f = ff;
    294 
    295   const FP32 f32infty = { 255 << 23 };
    296   const FP32 f16max = { (127 + 16) << 23 };
    297   const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
    298   unsigned int sign_mask = 0x80000000u;
    299   __half o;
    300   o.x = static_cast<unsigned short>(0x0u);
    301 
    302   unsigned int sign = f.u & sign_mask;
    303   f.u ^= sign;
    304 
    305   // NOTE all the integer compares in this function can be safely
    306   // compiled into signed compares since all operands are below
    307   // 0x80000000. Important if you want fast straight SSE2 code
    308   // (since there's no unsigned PCMPGTD).
    309 
    310   if (f.u >= f16max.u) {  // result is Inf or NaN (all exponent bits set)
    311     o.x = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
    312   } else {  // (De)normalized number or zero
    313     if (f.u < (113 << 23)) {  // resulting FP16 is subnormal or zero
    314       // use a magic value to align our 10 mantissa bits at the bottom of
    315       // the float. as long as FP addition is round-to-nearest-even this
    316       // just works.
    317       f.f += denorm_magic.f;
    318 
    319       // and one integer subtract of the bias later, we have our final float!
    320       o.x = static_cast<unsigned short>(f.u - denorm_magic.u);
    321     } else {
    322       unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
    323 
    324       // update exponent, rounding bias part 1
    325       f.u += ((unsigned int)(15 - 127) << 23) + 0xfff;
    326       // rounding bias part 2
    327       f.u += mant_odd;
    328       // take the bits!
    329       o.x = static_cast<unsigned short>(f.u >> 13);
    330     }
    331   }
    332 
    333   o.x |= static_cast<unsigned short>(sign >> 16);
    334   return o;
    335 #endif
    336 }
    337 
    338 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) {
    339 #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
    340   return __half2float(h);
    341 
    342 #elif defined(EIGEN_HAS_FP16_C)
    343   return _cvtsh_ss(h.x);
    344 
    345 #else
    346   const FP32 magic = { 113 << 23 };
    347   const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift
    348   FP32 o;
    349 
    350   o.u = (h.x & 0x7fff) << 13;             // exponent/mantissa bits
    351   unsigned int exp = shifted_exp & o.u;   // just the exponent
    352   o.u += (127 - 15) << 23;                // exponent adjust
    353 
    354   // handle exponent special cases
    355   if (exp == shifted_exp) {     // Inf/NaN?
    356     o.u += (128 - 16) << 23;    // extra exp adjust
    357   } else if (exp == 0) {        // Zero/Denormal?
    358     o.u += 1 << 23;             // extra exp adjust
    359     o.f -= magic.f;             // renormalize
    360   }
    361 
    362   o.u |= (h.x & 0x8000) << 16;    // sign bit
    363   return o.f;
    364 #endif
    365 }
    366 
    367 // --- standard functions ---
    368 
    369 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
    370   return (a.x & 0x7fff) == 0x7c00;
    371 }
    372 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
    373 #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
    374   return __hisnan(a);
    375 #else
    376   return (a.x & 0x7fff) > 0x7c00;
    377 #endif
    378 }
    379 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isfinite)(const half& a) {
    380   return !(isinf EIGEN_NOT_A_MACRO (a)) && !(isnan EIGEN_NOT_A_MACRO (a));
    381 }
    382 
    383 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
    384   half result;
    385   result.x = a.x & 0x7FFF;
    386   return result;
    387 }
    388 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
    389   return half(::expf(float(a)));
    390 }
    391 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
    392 #if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
    393   return Eigen::half(::hlog(a));
    394 #else
    395   return half(::logf(float(a)));
    396 #endif
    397 }
    398 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) {
    399   return half(numext::log1p(float(a)));
    400 }
    401 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
    402   return half(::log10f(float(a)));
    403 }
    404 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
    405   return half(::sqrtf(float(a)));
    406 }
    407 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) {
    408   return half(::powf(float(a), float(b)));
    409 }
    410 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sin(const half& a) {
    411   return half(::sinf(float(a)));
    412 }
    413 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half cos(const half& a) {
    414   return half(::cosf(float(a)));
    415 }
    416 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tan(const half& a) {
    417   return half(::tanf(float(a)));
    418 }
    419 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
    420   return half(::tanhf(float(a)));
    421 }
    422 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
    423   return half(::floorf(float(a)));
    424 }
    425 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
    426   return half(::ceilf(float(a)));
    427 }
    428 
    429 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
    430 #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
    431   return __hlt(b, a) ? b : a;
    432 #else
    433   const float f1 = static_cast<float>(a);
    434   const float f2 = static_cast<float>(b);
    435   return f2 < f1 ? b : a;
    436 #endif
    437 }
    438 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
    439 #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
    440   return __hlt(a, b) ? b : a;
    441 #else
    442   const float f1 = static_cast<float>(a);
    443   const float f2 = static_cast<float>(b);
    444   return f1 < f2 ? b : a;
    445 #endif
    446 }
    447 
    448 EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const half& v) {
    449   os << static_cast<float>(v);
    450   return os;
    451 }
    452 
    453 } // end namespace half_impl
    454 
    455 // import Eigen::half_impl::half into Eigen namespace
    456 // using half_impl::half;
    457 
    458 namespace internal {
    459 
    460 template<>
    461 struct random_default_impl<half, false, false>
    462 {
    463   static inline half run(const half& x, const half& y)
    464   {
    465     return x + (y-x) * half(float(std::rand()) / float(RAND_MAX));
    466   }
    467   static inline half run()
    468   {
    469     return run(half(-1.f), half(1.f));
    470   }
    471 };
    472 
    473 template<> struct is_arithmetic<half> { enum { value = true }; };
    474 
    475 } // end namespace internal
    476 
    477 }  // end namespace Eigen
    478 
    479 namespace std {
    480 template<>
    481 struct numeric_limits<Eigen::half> {
    482   static const bool is_specialized = true;
    483   static const bool is_signed = true;
    484   static const bool is_integer = false;
    485   static const bool is_exact = false;
    486   static const bool has_infinity = true;
    487   static const bool has_quiet_NaN = true;
    488   static const bool has_signaling_NaN = true;
    489   static const float_denorm_style has_denorm = denorm_present;
    490   static const bool has_denorm_loss = false;
    491   static const std::float_round_style round_style = std::round_to_nearest;
    492   static const bool is_iec559 = false;
    493   static const bool is_bounded = false;
    494   static const bool is_modulo = false;
    495   static const int digits = 11;
    496   static const int digits10 = 2;
    497   //static const int max_digits10 = ;
    498   static const int radix = 2;
    499   static const int min_exponent = -13;
    500   static const int min_exponent10 = -4;
    501   static const int max_exponent = 16;
    502   static const int max_exponent10 = 4;
    503   static const bool traps = true;
    504   static const bool tinyness_before = false;
    505 
    506   static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
    507   static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
    508   static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
    509   static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
    510   static Eigen::half round_error() { return Eigen::half(0.5); }
    511   static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
    512   static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
    513   static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
    514   static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
    515 };
    516 }
    517 
    518 namespace Eigen {
    519 
    520 template<> struct NumTraits<Eigen::half>
    521     : GenericNumTraits<Eigen::half>
    522 {
    523   enum {
    524     IsSigned = true,
    525     IsInteger = false,
    526     IsComplex = false,
    527     RequireInitialization = false
    528   };
    529 
    530   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() {
    531     return half_impl::raw_uint16_to_half(0x0800);
    532   }
    533   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half dummy_precision() { return Eigen::half(1e-2f); }
    534   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half highest() {
    535     return half_impl::raw_uint16_to_half(0x7bff);
    536   }
    537   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half lowest() {
    538     return half_impl::raw_uint16_to_half(0xfbff);
    539   }
    540   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half infinity() {
    541     return half_impl::raw_uint16_to_half(0x7c00);
    542   }
    543   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half quiet_NaN() {
    544     return half_impl::raw_uint16_to_half(0x7c01);
    545   }
    546 };
    547 
    548 } // end namespace Eigen
    549 
    550 // C-like standard mathematical functions and trancendentals.
    551 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half fabsh(const Eigen::half& a) {
    552   Eigen::half result;
    553   result.x = a.x & 0x7FFF;
    554   return result;
    555 }
    556 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
    557   return Eigen::half(::expf(float(a)));
    558 }
    559 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
    560 #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
    561   return Eigen::half(::hlog(a));
    562 #else
    563   return Eigen::half(::logf(float(a)));
    564 #endif
    565 }
    566 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half sqrth(const Eigen::half& a) {
    567   return Eigen::half(::sqrtf(float(a)));
    568 }
    569 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half powh(const Eigen::half& a, const Eigen::half& b) {
    570   return Eigen::half(::powf(float(a), float(b)));
    571 }
    572 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half floorh(const Eigen::half& a) {
    573   return Eigen::half(::floorf(float(a)));
    574 }
    575 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half ceilh(const Eigen::half& a) {
    576   return Eigen::half(::ceilf(float(a)));
    577 }
    578 
    579 namespace std {
    580 
    581 #if __cplusplus > 199711L
    582 template <>
    583 struct hash<Eigen::half> {
    584   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::half& a) const {
    585     return static_cast<std::size_t>(a.x);
    586   }
    587 };
    588 #endif
    589 
    590 } // end namespace std
    591 
    592 
    593 // Add the missing shfl_xor intrinsic
    594 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
    595 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
    596   return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
    597 }
    598 #endif
    599 
    600 // ldg() has an overload for __half, but we also need one for Eigen::half.
    601 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
    602 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
    603   return Eigen::half_impl::raw_uint16_to_half(
    604       __ldg(reinterpret_cast<const unsigned short*>(ptr)));
    605 }
    606 #endif
    607 
    608 
    609 #if defined(__CUDA_ARCH__)
    610 namespace Eigen {
    611 namespace numext {
    612 
    613 template<>
    614 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
    615 bool (isnan)(const Eigen::half& h) {
    616   return (half_impl::isnan)(h);
    617 }
    618 
    619 template<>
    620 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
    621 bool (isinf)(const Eigen::half& h) {
    622   return (half_impl::isinf)(h);
    623 }
    624 
    625 template<>
    626 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
    627 bool (isfinite)(const Eigen::half& h) {
    628   return (half_impl::isfinite)(h);
    629 }
    630 
    631 } // namespace Eigen
    632 }  // namespace numext
    633 #endif
    634 
    635 #endif // EIGEN_HALF_CUDA_H
    636