xref: /aosp_15_r20/external/eigen/Eigen/src/Core/arch/Default/Half.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
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 either from CUDA's or HIP'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_H
37 #define EIGEN_HALF_H
38 
39 #include <sstream>
40 
41 #if defined(EIGEN_HAS_GPU_FP16) || defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
42 // When compiling with GPU support, the "__half_raw" base class as well as
43 // some other routines are defined in the GPU compiler header files
44 // (cuda_fp16.h, hip_fp16.h), and they are not tagged constexpr
45 // As a consequence, we get compile failures when compiling Eigen with
46 // GPU support. Hence the need to disable EIGEN_CONSTEXPR when building
47 // Eigen with GPU support
48   #pragma push_macro("EIGEN_CONSTEXPR")
49   #undef EIGEN_CONSTEXPR
50   #define EIGEN_CONSTEXPR
51 #endif
52 
53 #define F16_PACKET_FUNCTION(PACKET_F, PACKET_F16, METHOD)           \
54   template <>                                                       \
55   EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC EIGEN_UNUSED                \
56   PACKET_F16 METHOD<PACKET_F16>(const PACKET_F16& _x) {             \
57     return float2half(METHOD<PACKET_F>(half2float(_x)));            \
58   }
59 
60 namespace Eigen {
61 
62 struct half;
63 
64 namespace half_impl {
65 
66 // We want to use the __half_raw struct from the HIP header file only during the device compile phase.
67 // This is required because of a quirk in the way TensorFlow GPU builds are done.
68 // When compiling TensorFlow source code with GPU support, files that
69 //  * contain GPU kernels (i.e. *.cu.cc files) are compiled via hipcc
70 //  * do not contain GPU kernels ( i.e. *.cc files) are compiled via gcc (typically)
71 //
72 // Tensorflow uses the Eigen::half type as its FP16 type, and there are functions that
73 //  * are defined in a file that gets compiled via hipcc AND
74 //  * have Eigen::half as a pass-by-value argument AND
75 //  * are called in a file that gets compiled via gcc
76 //
77 // In the scenario described above the caller and callee will see different versions
78 // of the Eigen::half base class __half_raw, and they will be compiled by different compilers
79 //
80 // There appears to be an ABI mismatch between gcc and clang (which is called by hipcc) that results in
81 // the callee getting corrupted values for the Eigen::half argument.
82 //
83 // Making the host side compile phase of hipcc use the same Eigen::half impl, as the gcc compile, resolves
84 // this error, and hence the following convoluted #if condition
85 #if !defined(EIGEN_HAS_GPU_FP16) || !defined(EIGEN_GPU_COMPILE_PHASE)
86 // Make our own __half_raw definition that is similar to CUDA's.
87 struct __half_raw {
88 #if (defined(EIGEN_HAS_GPU_FP16) && !defined(EIGEN_GPU_COMPILE_PHASE))
89   // Eigen::half can be used as the datatype for shared memory declarations (in Eigen and TF)
90   // The element type for shared memory cannot have non-trivial constructors
91   // and hence the following special casing (which skips the zero-initilization).
92   // Note that this check gets done even in the host compilation phase, and
93   // hence the need for this
__half_raw__half_raw94   EIGEN_DEVICE_FUNC __half_raw() {}
95 #else
96   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __half_raw() : x(0) {}
97 #endif
98 #if defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
__half_raw__half_raw99   explicit EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __half_raw(numext::uint16_t raw) : x(numext::bit_cast<__fp16>(raw)) {
100   }
101   __fp16 x;
102 #else
__half_raw__half_raw103   explicit EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __half_raw(numext::uint16_t raw) : x(raw) {}
104   numext::uint16_t x;
105 #endif
106 };
107 
108 #elif defined(EIGEN_HAS_HIP_FP16)
109   // Nothing to do here
110   // HIP fp16 header file has a definition for __half_raw
111 #elif defined(EIGEN_HAS_CUDA_FP16)
112   #if EIGEN_CUDA_SDK_VER < 90000
113     // In CUDA < 9.0, __half is the equivalent of CUDA 9's __half_raw
114     typedef __half __half_raw;
115   #endif // defined(EIGEN_HAS_CUDA_FP16)
116 #elif defined(SYCL_DEVICE_ONLY)
117   typedef cl::sycl::half __half_raw;
118 #endif
119 
120 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __half_raw raw_uint16_to_half(numext::uint16_t x);
121 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff);
122 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h);
123 
124 struct half_base : public __half_raw {
half_basehalf_base125   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half_base() {}
half_basehalf_base126   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half_base(const __half_raw& h) : __half_raw(h) {}
127 
128 #if defined(EIGEN_HAS_GPU_FP16)
129  #if defined(EIGEN_HAS_HIP_FP16)
half_basehalf_base130   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half_base(const __half& h) { x = __half_as_ushort(h); }
131  #elif defined(EIGEN_HAS_CUDA_FP16)
132   #if EIGEN_CUDA_SDK_VER >= 90000
half_basehalf_base133   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half_base(const __half& h) : __half_raw(*(__half_raw*)&h) {}
134   #endif
135  #endif
136 #endif
137 };
138 
139 } // namespace half_impl
140 
141 // Class definition.
142 struct half : public half_impl::half_base {
143 
144   // Writing this out as separate #if-else blocks to make the code easier to follow
145   // The same applies to most #if-else blocks in this file
146 #if !defined(EIGEN_HAS_GPU_FP16) || !defined(EIGEN_GPU_COMPILE_PHASE)
147   // Use the same base class for the following two scenarios
148   // * when compiling without GPU support enabled
149   // * during host compile phase when compiling with GPU support enabled
150   typedef half_impl::__half_raw __half_raw;
151 #elif defined(EIGEN_HAS_HIP_FP16)
152   // Nothing to do here
153   // HIP fp16 header file has a definition for __half_raw
154 #elif defined(EIGEN_HAS_CUDA_FP16)
155   // Note that EIGEN_CUDA_SDK_VER is set to 0 even when compiling with HIP, so
156   // (EIGEN_CUDA_SDK_VER < 90000) is true even for HIP!  So keeping this within
157   // #if defined(EIGEN_HAS_CUDA_FP16) is needed
158   #if defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
159     typedef half_impl::__half_raw __half_raw;
160   #endif
161 #endif
162 
halfhalf163   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half() {}
164 
halfhalf165   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half(const __half_raw& h) : half_impl::half_base(h) {}
166 
167 #if defined(EIGEN_HAS_GPU_FP16)
168  #if defined(EIGEN_HAS_HIP_FP16)
halfhalf169   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half(const __half& h) : half_impl::half_base(h) {}
170  #elif defined(EIGEN_HAS_CUDA_FP16)
171   #if defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER >= 90000
halfhalf172   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half(const __half& h) : half_impl::half_base(h) {}
173   #endif
174  #endif
175 #endif
176 
177 
halfhalf178   explicit EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR half(bool b)
179       : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
180   template<class T>
halfhalf181   explicit EIGEN_DEVICE_FUNC half(T val)
182       : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(val))) {}
halfhalf183   explicit EIGEN_DEVICE_FUNC half(float f)
184       : half_impl::half_base(half_impl::float_to_half_rtne(f)) {}
185 
186   // Following the convention of numpy, converting between complex and
187   // float will lead to loss of imag value.
188   template<typename RealScalar>
halfhalf189   explicit EIGEN_DEVICE_FUNC half(std::complex<RealScalar> c)
190       : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(c.real()))) {}
191 
192    EIGEN_DEVICE_FUNC operator float() const {  // NOLINT: Allow implicit conversion to float, because it is lossless.
193     return half_impl::half_to_float(*this);
194   }
195 
196 #if defined(EIGEN_HAS_GPU_FP16) && !defined(EIGEN_GPU_COMPILE_PHASE)
__halfhalf197   EIGEN_DEVICE_FUNC operator __half() const {
198     ::__half_raw hr;
199     hr.x = x;
200     return __half(hr);
201   }
202 #endif
203 };
204 
205 } // end namespace Eigen
206 
207 namespace std {
208 template<>
209 struct numeric_limits<Eigen::half> {
210   static const bool is_specialized = true;
211   static const bool is_signed = true;
212   static const bool is_integer = false;
213   static const bool is_exact = false;
214   static const bool has_infinity = true;
215   static const bool has_quiet_NaN = true;
216   static const bool has_signaling_NaN = true;
217   static const float_denorm_style has_denorm = denorm_present;
218   static const bool has_denorm_loss = false;
219   static const std::float_round_style round_style = std::round_to_nearest;
220   static const bool is_iec559 = false;
221   static const bool is_bounded = false;
222   static const bool is_modulo = false;
223   static const int digits = 11;
224   static const int digits10 = 3;      // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
225   static const int max_digits10 = 5;  // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
226   static const int radix = 2;
227   static const int min_exponent = -13;
228   static const int min_exponent10 = -4;
229   static const int max_exponent = 16;
230   static const int max_exponent10 = 4;
231   static const bool traps = true;
232   static const bool tinyness_before = false;
233 
234   static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
235   static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
236   static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
237   static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
238   static Eigen::half round_error() { return Eigen::half(0.5); }
239   static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
240   static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
241   static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7d00); }
242   static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
243 };
244 
245 // If std::numeric_limits<T> is specialized, should also specialize
246 // std::numeric_limits<const T>, std::numeric_limits<volatile T>, and
247 // std::numeric_limits<const volatile T>
248 // https://stackoverflow.com/a/16519653/
249 template<>
250 struct numeric_limits<const Eigen::half> : numeric_limits<Eigen::half> {};
251 template<>
252 struct numeric_limits<volatile Eigen::half> : numeric_limits<Eigen::half> {};
253 template<>
254 struct numeric_limits<const volatile Eigen::half> : numeric_limits<Eigen::half> {};
255 } // end namespace std
256 
257 namespace Eigen {
258 
259 namespace half_impl {
260 
261 #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && \
262      EIGEN_CUDA_ARCH >= 530) ||                                  \
263     (defined(EIGEN_HAS_HIP_FP16) && defined(HIP_DEVICE_COMPILE))
264 // Note: We deliberatly do *not* define this to 1 even if we have Arm's native
265 // fp16 type since GPU halfs are rather different from native CPU halfs.
266 // TODO: Rename to something like EIGEN_HAS_NATIVE_GPU_FP16
267 #define EIGEN_HAS_NATIVE_FP16
268 #endif
269 
270 // Intrinsics for native fp16 support. Note that on current hardware,
271 // these are no faster than fp32 arithmetic (you need to use the half2
272 // versions to get the ALU speed increased), but you do save the
273 // conversion steps back and forth.
274 
275 #if defined(EIGEN_HAS_NATIVE_FP16)
276 EIGEN_STRONG_INLINE __device__ half operator + (const half& a, const half& b) {
277 #if defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER >= 90000
278   return __hadd(::__half(a), ::__half(b));
279 #else
280   return __hadd(a, b);
281 #endif
282 }
283 EIGEN_STRONG_INLINE __device__ half operator * (const half& a, const half& b) {
284   return __hmul(a, b);
285 }
286 EIGEN_STRONG_INLINE __device__ half operator - (const half& a, const half& b) {
287   return __hsub(a, b);
288 }
289 EIGEN_STRONG_INLINE __device__ half operator / (const half& a, const half& b) {
290 #if defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER >= 90000
291   return __hdiv(a, b);
292 #else
293   float num = __half2float(a);
294   float denom = __half2float(b);
295   return __float2half(num / denom);
296 #endif
297 }
298 EIGEN_STRONG_INLINE __device__ half operator - (const half& a) {
299   return __hneg(a);
300 }
301 EIGEN_STRONG_INLINE __device__ half& operator += (half& a, const half& b) {
302   a = a + b;
303   return a;
304 }
305 EIGEN_STRONG_INLINE __device__ half& operator *= (half& a, const half& b) {
306   a = a * b;
307   return a;
308 }
309 EIGEN_STRONG_INLINE __device__ half& operator -= (half& a, const half& b) {
310   a = a - b;
311   return a;
312 }
313 EIGEN_STRONG_INLINE __device__ half& operator /= (half& a, const half& b) {
314   a = a / b;
315   return a;
316 }
317 EIGEN_STRONG_INLINE __device__ bool operator == (const half& a, const half& b) {
318   return __heq(a, b);
319 }
320 EIGEN_STRONG_INLINE __device__ bool operator != (const half& a, const half& b) {
321   return __hne(a, b);
322 }
323 EIGEN_STRONG_INLINE __device__ bool operator < (const half& a, const half& b) {
324   return __hlt(a, b);
325 }
326 EIGEN_STRONG_INLINE __device__ bool operator <= (const half& a, const half& b) {
327   return __hle(a, b);
328 }
329 EIGEN_STRONG_INLINE __device__ bool operator > (const half& a, const half& b) {
330   return __hgt(a, b);
331 }
332 EIGEN_STRONG_INLINE __device__ bool operator >= (const half& a, const half& b) {
333   return __hge(a, b);
334 }
335 #endif
336 
337 #if defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
338 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
339   return half(vaddh_f16(a.x, b.x));
340 }
341 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
342   return half(vmulh_f16(a.x, b.x));
343 }
344 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
345   return half(vsubh_f16(a.x, b.x));
346 }
347 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
348   return half(vdivh_f16(a.x, b.x));
349 }
350 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
351   return half(vnegh_f16(a.x));
352 }
353 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
354   a = half(vaddh_f16(a.x, b.x));
355   return a;
356 }
357 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
358   a = half(vmulh_f16(a.x, b.x));
359   return a;
360 }
361 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
362   a = half(vsubh_f16(a.x, b.x));
363   return a;
364 }
365 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
366   a = half(vdivh_f16(a.x, b.x));
367   return a;
368 }
369 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
370   return vceqh_f16(a.x, b.x);
371 }
372 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
373   return !vceqh_f16(a.x, b.x);
374 }
375 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
376   return vclth_f16(a.x, b.x);
377 }
378 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
379   return vcleh_f16(a.x, b.x);
380 }
381 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
382   return vcgth_f16(a.x, b.x);
383 }
384 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
385   return vcgeh_f16(a.x, b.x);
386 }
387 // We need to distinguish ‘clang as the CUDA compiler’ from ‘clang as the host compiler,
388 // invoked by NVCC’ (e.g. on MacOS). The former needs to see both host and device implementation
389 // of the functions, while the latter can only deal with one of them.
390 #elif !defined(EIGEN_HAS_NATIVE_FP16) || (EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC) // Emulate support for half floats
391 
392 #if EIGEN_COMP_CLANG && defined(EIGEN_CUDACC)
393 // We need to provide emulated *host-side* FP16 operators for clang.
394 #pragma push_macro("EIGEN_DEVICE_FUNC")
395 #undef EIGEN_DEVICE_FUNC
396 #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_HAS_NATIVE_FP16)
397 #define EIGEN_DEVICE_FUNC __host__
398 #else // both host and device need emulated ops.
399 #define EIGEN_DEVICE_FUNC __host__ __device__
400 #endif
401 #endif
402 
403 // Definitions for CPUs and older HIP+CUDA, mostly working through conversion
404 // to/from fp32.
405 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
406   return half(float(a) + float(b));
407 }
408 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
409   return half(float(a) * float(b));
410 }
411 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
412   return half(float(a) - float(b));
413 }
414 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
415   return half(float(a) / float(b));
416 }
417 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
418   half result;
419   result.x = a.x ^ 0x8000;
420   return result;
421 }
422 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
423   a = half(float(a) + float(b));
424   return a;
425 }
426 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
427   a = half(float(a) * float(b));
428   return a;
429 }
430 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
431   a = half(float(a) - float(b));
432   return a;
433 }
434 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
435   a = half(float(a) / float(b));
436   return a;
437 }
438 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
439   return numext::equal_strict(float(a),float(b));
440 }
441 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
442   return numext::not_equal_strict(float(a), float(b));
443 }
444 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
445   return float(a) < float(b);
446 }
447 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
448   return float(a) <= float(b);
449 }
450 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
451   return float(a) > float(b);
452 }
453 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
454   return float(a) >= float(b);
455 }
456 
457 #if defined(__clang__) && defined(__CUDA__)
458 #pragma pop_macro("EIGEN_DEVICE_FUNC")
459 #endif
460 #endif  // Emulate support for half floats
461 
462 // Division by an index. Do it in full float precision to avoid accuracy
463 // issues in converting the denominator to half.
464 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
465   return half(static_cast<float>(a) / static_cast<float>(b));
466 }
467 
468 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator++(half& a) {
469   a += half(1);
470   return a;
471 }
472 
473 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator--(half& a) {
474   a -= half(1);
475   return a;
476 }
477 
478 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator++(half& a, int) {
479   half original_value = a;
480   ++a;
481   return original_value;
482 }
483 
484 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator--(half& a, int) {
485   half original_value = a;
486   --a;
487   return original_value;
488 }
489 
490 // Conversion routines, including fallbacks for the host or older CUDA.
491 // Note that newer Intel CPUs (Haswell or newer) have vectorized versions of
492 // these in hardware. If we need more performance on older/other CPUs, they are
493 // also possible to vectorize directly.
494 
495 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR __half_raw raw_uint16_to_half(numext::uint16_t x) {
496   // We cannot simply do a "return __half_raw(x)" here, because __half_raw is union type
497   // in the hip_fp16 header file, and that will trigger a compile error
498   // On the other hand, having anything but a return statement also triggers a compile error
499   // because this is constexpr function.
500   // Fortunately, since we need to disable EIGEN_CONSTEXPR for GPU anyway, we can get out
501   // of this catch22 by having separate bodies for GPU / non GPU
502 #if defined(EIGEN_HAS_GPU_FP16)
503    __half_raw h;
504    h.x = x;
505   return h;
506 #else
507   return __half_raw(x);
508 #endif
509 }
510 
511 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC numext::uint16_t raw_half_as_uint16(const __half_raw& h) {
512   // HIP/CUDA/Default have a member 'x' of type uint16_t.
513   // For ARM64 native half, the member 'x' is of type __fp16, so we need to bit-cast.
514   // For SYCL, cl::sycl::half is _Float16, so cast directly.
515 #if defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
516   return numext::bit_cast<numext::uint16_t>(h.x);
517 #elif defined(SYCL_DEVICE_ONLY)
518   return numext::bit_cast<numext::uint16_t>(h);
519 #else
520   return h.x;
521 #endif
522 }
523 
524 union float32_bits {
525   unsigned int u;
526   float f;
527 };
528 
529 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
530 #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
531   (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
532   __half tmp_ff = __float2half(ff);
533   return *(__half_raw*)&tmp_ff;
534 
535 #elif defined(EIGEN_HAS_FP16_C)
536   __half_raw h;
537   h.x = _cvtss_sh(ff, 0);
538   return h;
539 
540 #elif defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
541   __half_raw h;
542   h.x = static_cast<__fp16>(ff);
543   return h;
544 
545 #else
546   float32_bits f; f.f = ff;
547 
548   const float32_bits f32infty = { 255 << 23 };
549   const float32_bits f16max = { (127 + 16) << 23 };
550   const float32_bits denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
551   unsigned int sign_mask = 0x80000000u;
552   __half_raw o;
553   o.x = static_cast<numext::uint16_t>(0x0u);
554 
555   unsigned int sign = f.u & sign_mask;
556   f.u ^= sign;
557 
558   // NOTE all the integer compares in this function can be safely
559   // compiled into signed compares since all operands are below
560   // 0x80000000. Important if you want fast straight SSE2 code
561   // (since there's no unsigned PCMPGTD).
562 
563   if (f.u >= f16max.u) {  // result is Inf or NaN (all exponent bits set)
564     o.x = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
565   } else {  // (De)normalized number or zero
566     if (f.u < (113 << 23)) {  // resulting FP16 is subnormal or zero
567       // use a magic value to align our 10 mantissa bits at the bottom of
568       // the float. as long as FP addition is round-to-nearest-even this
569       // just works.
570       f.f += denorm_magic.f;
571 
572       // and one integer subtract of the bias later, we have our final float!
573       o.x = static_cast<numext::uint16_t>(f.u - denorm_magic.u);
574     } else {
575       unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
576 
577       // update exponent, rounding bias part 1
578       // Equivalent to `f.u += ((unsigned int)(15 - 127) << 23) + 0xfff`, but
579       // without arithmetic overflow.
580       f.u += 0xc8000fffU;
581       // rounding bias part 2
582       f.u += mant_odd;
583       // take the bits!
584       o.x = static_cast<numext::uint16_t>(f.u >> 13);
585     }
586   }
587 
588   o.x |= static_cast<numext::uint16_t>(sign >> 16);
589   return o;
590 #endif
591 }
592 
593 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h) {
594 #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
595   (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
596   return __half2float(h);
597 #elif defined(EIGEN_HAS_FP16_C)
598   return _cvtsh_ss(h.x);
599 #elif defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
600   return static_cast<float>(h.x);
601 #else
602   const float32_bits magic = { 113 << 23 };
603   const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift
604   float32_bits o;
605 
606   o.u = (h.x & 0x7fff) << 13;             // exponent/mantissa bits
607   unsigned int exp = shifted_exp & o.u;   // just the exponent
608   o.u += (127 - 15) << 23;                // exponent adjust
609 
610   // handle exponent special cases
611   if (exp == shifted_exp) {     // Inf/NaN?
612     o.u += (128 - 16) << 23;    // extra exp adjust
613   } else if (exp == 0) {        // Zero/Denormal?
614     o.u += 1 << 23;             // extra exp adjust
615     o.f -= magic.f;             // renormalize
616   }
617 
618   o.u |= (h.x & 0x8000) << 16;    // sign bit
619   return o.f;
620 #endif
621 }
622 
623 // --- standard functions ---
624 
625 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
626 #ifdef EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC
627   return (numext::bit_cast<numext::uint16_t>(a.x) & 0x7fff) == 0x7c00;
628 #else
629   return (a.x & 0x7fff) == 0x7c00;
630 #endif
631 }
632 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
633 #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
634   (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
635   return __hisnan(a);
636 #elif defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
637   return (numext::bit_cast<numext::uint16_t>(a.x) & 0x7fff) > 0x7c00;
638 #else
639   return (a.x & 0x7fff) > 0x7c00;
640 #endif
641 }
642 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isfinite)(const half& a) {
643   return !(isinf EIGEN_NOT_A_MACRO (a)) && !(isnan EIGEN_NOT_A_MACRO (a));
644 }
645 
646 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
647 #if defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
648   return half(vabsh_f16(a.x));
649 #else
650   half result;
651   result.x = a.x & 0x7FFF;
652   return result;
653 #endif
654 }
655 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
656 #if (EIGEN_CUDA_SDK_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530) || \
657   defined(EIGEN_HIP_DEVICE_COMPILE)
658   return half(hexp(a));
659 #else
660    return half(::expf(float(a)));
661 #endif
662 }
663 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half expm1(const half& a) {
664   return half(numext::expm1(float(a)));
665 }
666 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
667 #if (defined(EIGEN_HAS_CUDA_FP16) && EIGEN_CUDA_SDK_VER >= 80000 && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
668   (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
669   return half(::hlog(a));
670 #else
671   return half(::logf(float(a)));
672 #endif
673 }
674 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) {
675   return half(numext::log1p(float(a)));
676 }
677 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
678   return half(::log10f(float(a)));
679 }
680 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log2(const half& a) {
681   return half(static_cast<float>(EIGEN_LOG2E) * ::logf(float(a)));
682 }
683 
684 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
685 #if (EIGEN_CUDA_SDK_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530) || \
686   defined(EIGEN_HIP_DEVICE_COMPILE)
687   return half(hsqrt(a));
688 #else
689     return half(::sqrtf(float(a)));
690 #endif
691 }
692 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) {
693   return half(::powf(float(a), float(b)));
694 }
695 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sin(const half& a) {
696   return half(::sinf(float(a)));
697 }
698 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half cos(const half& a) {
699   return half(::cosf(float(a)));
700 }
701 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tan(const half& a) {
702   return half(::tanf(float(a)));
703 }
704 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
705   return half(::tanhf(float(a)));
706 }
707 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half asin(const half& a) {
708   return half(::asinf(float(a)));
709 }
710 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half acos(const half& a) {
711   return half(::acosf(float(a)));
712 }
713 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
714 #if (EIGEN_CUDA_SDK_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300) || \
715   defined(EIGEN_HIP_DEVICE_COMPILE)
716   return half(hfloor(a));
717 #else
718   return half(::floorf(float(a)));
719 #endif
720 }
721 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
722 #if (EIGEN_CUDA_SDK_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300) || \
723   defined(EIGEN_HIP_DEVICE_COMPILE)
724   return half(hceil(a));
725 #else
726   return half(::ceilf(float(a)));
727 #endif
728 }
729 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half rint(const half& a) {
730   return half(::rintf(float(a)));
731 }
732 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half round(const half& a) {
733   return half(::roundf(float(a)));
734 }
735 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half fmod(const half& a, const half& b) {
736   return half(::fmodf(float(a), float(b)));
737 }
738 
739 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
740 #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
741   (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
742   return __hlt(b, a) ? b : a;
743 #else
744   const float f1 = static_cast<float>(a);
745   const float f2 = static_cast<float>(b);
746   return f2 < f1 ? b : a;
747 #endif
748 }
749 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
750 #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 530) || \
751   (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
752   return __hlt(a, b) ? b : a;
753 #else
754   const float f1 = static_cast<float>(a);
755   const float f2 = static_cast<float>(b);
756   return f1 < f2 ? b : a;
757 #endif
758 }
759 
760 #ifndef EIGEN_NO_IO
761 EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const half& v) {
762   os << static_cast<float>(v);
763   return os;
764 }
765 #endif
766 
767 } // end namespace half_impl
768 
769 // import Eigen::half_impl::half into Eigen namespace
770 // using half_impl::half;
771 
772 namespace internal {
773 
774 template<>
775 struct random_default_impl<half, false, false>
776 {
777   static inline half run(const half& x, const half& y)
778   {
779     return x + (y-x) * half(float(std::rand()) / float(RAND_MAX));
780   }
781   static inline half run()
782   {
783     return run(half(-1.f), half(1.f));
784   }
785 };
786 
787 template<> struct is_arithmetic<half> { enum { value = true }; };
788 
789 } // end namespace internal
790 
791 template<> struct NumTraits<Eigen::half>
792     : GenericNumTraits<Eigen::half>
793 {
794   enum {
795     IsSigned = true,
796     IsInteger = false,
797     IsComplex = false,
798     RequireInitialization = false
799   };
800 
801   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::half epsilon() {
802     return half_impl::raw_uint16_to_half(0x0800);
803   }
804   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::half dummy_precision() {
805     return half_impl::raw_uint16_to_half(0x211f); //  Eigen::half(1e-2f);
806   }
807   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::half highest() {
808     return half_impl::raw_uint16_to_half(0x7bff);
809   }
810   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::half lowest() {
811     return half_impl::raw_uint16_to_half(0xfbff);
812   }
813   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::half infinity() {
814     return half_impl::raw_uint16_to_half(0x7c00);
815   }
816   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR static EIGEN_STRONG_INLINE Eigen::half quiet_NaN() {
817     return half_impl::raw_uint16_to_half(0x7e00);
818   }
819 };
820 
821 } // end namespace Eigen
822 
823 #if defined(EIGEN_HAS_GPU_FP16) || defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
824   #pragma pop_macro("EIGEN_CONSTEXPR")
825 #endif
826 
827 namespace Eigen {
828 namespace numext {
829 
830 #if defined(EIGEN_GPU_COMPILE_PHASE)
831 
832 template <>
833 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool(isnan)(const Eigen::half& h) {
834   return (half_impl::isnan)(h);
835 }
836 
837 template <>
838 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool(isinf)(const Eigen::half& h) {
839   return (half_impl::isinf)(h);
840 }
841 
842 template <>
843 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE bool(isfinite)(const Eigen::half& h) {
844   return (half_impl::isfinite)(h);
845 }
846 
847 #endif
848 
849 template <>
850 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half bit_cast<Eigen::half, uint16_t>(const uint16_t& src) {
851   return Eigen::half(Eigen::half_impl::raw_uint16_to_half(src));
852 }
853 
854 template <>
855 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC uint16_t bit_cast<uint16_t, Eigen::half>(const Eigen::half& src) {
856   return Eigen::half_impl::raw_half_as_uint16(src);
857 }
858 
859 }  // namespace numext
860 }  // namespace Eigen
861 
862 // Add the missing shfl* intrinsics.
863 // The __shfl* functions are only valid on HIP or _CUDA_ARCH_ >= 300.
864 //   CUDA defines them for (__CUDA_ARCH__ >= 300 || !defined(__CUDA_ARCH__))
865 //
866 // HIP and CUDA prior to SDK 9.0 define
867 //    __shfl, __shfl_up, __shfl_down, __shfl_xor for int and float
868 // CUDA since 9.0 deprecates those and instead defines
869 //    __shfl_sync, __shfl_up_sync, __shfl_down_sync, __shfl_xor_sync,
870 //    with native support for __half and __nv_bfloat16
871 //
872 // Note that the following are __device__ - only functions.
873 #if (defined(EIGEN_CUDACC) && (!defined(EIGEN_CUDA_ARCH) || EIGEN_CUDA_ARCH >= 300)) \
874     || defined(EIGEN_HIPCC)
875 
876 #if defined(EIGEN_HAS_CUDA_FP16) && EIGEN_CUDA_SDK_VER >= 90000
877 
878 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_sync(unsigned mask, Eigen::half var, int srcLane, int width=warpSize) {
879   const __half h = var;
880   return static_cast<Eigen::half>(__shfl_sync(mask, h, srcLane, width));
881 }
882 
883 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_up_sync(unsigned mask, Eigen::half var, unsigned int delta, int width=warpSize) {
884   const __half h = var;
885   return static_cast<Eigen::half>(__shfl_up_sync(mask, h, delta, width));
886 }
887 
888 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_down_sync(unsigned mask, Eigen::half var, unsigned int delta, int width=warpSize) {
889   const __half h = var;
890   return static_cast<Eigen::half>(__shfl_down_sync(mask, h, delta, width));
891 }
892 
893 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor_sync(unsigned mask, Eigen::half var, int laneMask, int width=warpSize) {
894   const __half h = var;
895   return static_cast<Eigen::half>(__shfl_xor_sync(mask, h, laneMask, width));
896 }
897 
898 #else // HIP or CUDA SDK < 9.0
899 
900 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl(Eigen::half var, int srcLane, int width=warpSize) {
901   const int ivar = static_cast<int>(Eigen::numext::bit_cast<Eigen::numext::uint16_t>(var));
902   return Eigen::numext::bit_cast<Eigen::half>(static_cast<Eigen::numext::uint16_t>(__shfl(ivar, srcLane, width)));
903 }
904 
905 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_up(Eigen::half var, unsigned int delta, int width=warpSize) {
906   const int ivar = static_cast<int>(Eigen::numext::bit_cast<Eigen::numext::uint16_t>(var));
907   return Eigen::numext::bit_cast<Eigen::half>(static_cast<Eigen::numext::uint16_t>(__shfl_up(ivar, delta, width)));
908 }
909 
910 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_down(Eigen::half var, unsigned int delta, int width=warpSize) {
911   const int ivar = static_cast<int>(Eigen::numext::bit_cast<Eigen::numext::uint16_t>(var));
912   return Eigen::numext::bit_cast<Eigen::half>(static_cast<Eigen::numext::uint16_t>(__shfl_down(ivar, delta, width)));
913 }
914 
915 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
916   const int ivar = static_cast<int>(Eigen::numext::bit_cast<Eigen::numext::uint16_t>(var));
917   return Eigen::numext::bit_cast<Eigen::half>(static_cast<Eigen::numext::uint16_t>(__shfl_xor(ivar, laneMask, width)));
918 }
919 
920 #endif // HIP vs CUDA
921 #endif // __shfl*
922 
923 // ldg() has an overload for __half_raw, but we also need one for Eigen::half.
924 #if (defined(EIGEN_CUDACC) && (!defined(EIGEN_CUDA_ARCH) || EIGEN_CUDA_ARCH >= 350)) \
925     || defined(EIGEN_HIPCC)
926 EIGEN_STRONG_INLINE __device__ Eigen::half __ldg(const Eigen::half* ptr) {
927   return Eigen::half_impl::raw_uint16_to_half(__ldg(reinterpret_cast<const Eigen::numext::uint16_t*>(ptr)));
928 }
929 #endif // __ldg
930 
931 #if EIGEN_HAS_STD_HASH
932 namespace std {
933 template <>
934 struct hash<Eigen::half> {
935   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::half& a) const {
936     return static_cast<std::size_t>(Eigen::numext::bit_cast<Eigen::numext::uint16_t>(a));
937   }
938 };
939 } // end namespace std
940 #endif
941 
942 #endif // EIGEN_HALF_H
943