26 #ifndef MSHADOW_HALF_H_ 27 #define MSHADOW_HALF_H_ 31 #include <x86intrin.h> 32 #endif // MSHADOW_USE_F16C 36 #ifndef MSHADOW_HALF_ROUND_TO_NEAREST 37 #define MSHADOW_HALF_ROUND_TO_NEAREST 1 40 #if (MSHADOW_USE_CUDA && CUDA_VERSION >= 7050) 41 #define MSHADOW_CUDA_HALF 1 42 #include <cuda_fp16.h> 43 #if defined(__CUDA_ARCH__) 45 __host__ __device__
float __half2float_warp(
const volatile __half& h) {
47 #if CUDA_VERSION >= 9000 48 val =
const_cast<__half&
>(h);
52 return __half2float(val);
56 #define MSHADOW_CUDA_HALF 0 63 #define MSHADOW_HALF_OPERATOR(RTYPE, OP) \ 64 MSHADOW_XINLINE RTYPE operator OP (half_t a, half_t b) { \ 65 return RTYPE(float(a) OP float(b)); \ 67 template<typename T> \ 68 MSHADOW_XINLINE RTYPE operator OP (half_t a, T b) { \ 69 return RTYPE(float(a) OP float(b)); \ 71 template<typename T> \ 72 MSHADOW_XINLINE RTYPE operator OP (T a, half_t b) { \ 73 return RTYPE(float(a) OP float(b)); \ 76 #define MSHADOW_HALF_ASSIGNOP(AOP, OP) \ 77 template<typename T> \ 78 MSHADOW_XINLINE half_t operator AOP (const T& a) { \ 79 return *this = half_t(float(*this) OP float(a)); \ 81 template<typename T> \ 82 MSHADOW_XINLINE half_t operator AOP (const volatile T& a) volatile { \ 83 return *this = half_t(float(*this) OP float(a)); \ 86 #if (MSHADOW_CUDA_HALF && defined(__CUDA_ARCH__)) 87 #define MSHADOW_HALF_CONVERSIONOP(T) \ 88 MSHADOW_XINLINE operator T() const { \ 89 return T(__half2float(cuhalf_)); \ 91 MSHADOW_XINLINE operator T() const volatile { \ 92 return T(__half2float_warp(cuhalf_)); \ 94 #elif(MSHADOW_USE_F16C) 95 #define MSHADOW_HALF_CONVERSIONOP(T) \ 96 MSHADOW_XINLINE operator T() const { \ 97 return T(_cvtsh_ss(half_)); \ 99 MSHADOW_XINLINE operator T() const volatile { \ 100 return T(_cvtsh_ss(half_)); \ 103 #define MSHADOW_HALF_CONVERSIONOP(T) \ 104 MSHADOW_XINLINE operator T() const { \ 105 return T(half2float(half_)); \ 107 MSHADOW_XINLINE operator T() const volatile { \ 108 return T(half2float(half_)); \ 110 #endif // (MSHADOW_CUDA_HALF && defined(__CUDA_ARCH__)) 116 #if MSHADOW_CUDA_HALF 118 #endif // MSHADOW_CUDA_HALF 129 #if MSHADOW_CUDA_HALF 133 #endif // MSHADOW_CUDA_HALF 136 MSHADOW_XINLINE explicit half_t(
const double& value) { constructor(value); }
137 MSHADOW_XINLINE explicit half_t(
const int8_t& value) { constructor(value); }
138 MSHADOW_XINLINE explicit half_t(
const uint8_t& value) { constructor(value); }
139 MSHADOW_XINLINE explicit half_t(
const int32_t& value) { constructor(value); }
140 MSHADOW_XINLINE explicit half_t(
const uint32_t& value) { constructor(value); }
141 MSHADOW_XINLINE explicit half_t(
const int64_t& value) { constructor(value); }
142 MSHADOW_XINLINE explicit half_t(
const uint64_t& value) { constructor(value); }
156 return half_t(-
float(*
this));
166 return *
this = half_t(a);
176 return *
this = half_t(a);
186 static int const fp16FractionBits = 10;
187 static int const fp32FractionBits = 23;
188 static int32_t
const fp32FractionMask = ~(~0u << fp32FractionBits);
189 static int32_t
const fp32HiddenBit = 1 << fp32FractionBits;
190 static int const shift = fp32FractionBits - fp16FractionBits;
191 static int const shiftSign = 16;
192 static int32_t
const expAdjust = 127 - 15;
194 static int32_t
const infN = 0x7F800000;
195 static int32_t
const maxN = 0x477FFFFF;
196 static int32_t
const minN = 0x38800000;
197 static int32_t
const maxZ = 0x33000000;
198 static int32_t
const signN = 0x80000000;
200 static int32_t
const infC = infN >> shift;
201 static int32_t
const nanN = (infC + 1) << shift;
202 static int32_t
const maxC = maxN >> shift;
203 static int32_t
const minC = minN >> shift;
204 static int32_t
const signC = signN >> shiftSign;
206 static int32_t
const mulN = 0x52000000;
207 static int32_t
const mulC = 0x33800000;
209 static int32_t
const subC = 0x003FF;
210 static int32_t
const norC = 0x00400;
212 static int32_t
const maxD = infC - maxC - 1;
213 static int32_t
const minD = minC - subC - 1;
218 uint32_t sign = v.si & signN;
225 }
else if (v.si < minN) {
227 uint32_t exp32 = v.ui >> fp32FractionBits;
228 int32_t exp16 = exp32 - expAdjust;
231 uint32_t vshift = 1 - exp16;
232 uint32_t significand = fp32HiddenBit | (v.ui & fp32FractionMask);
233 v.ui = significand >> vshift;
241 #if MSHADOW_HALF_ROUND_TO_NEAREST == 1 243 v.ui += (v.ui & 0x3fff) != 0x1000 || (significand & 0x7ff) ? 0x1000 : 0;
245 }
else if (v.si <= maxN) {
247 #if MSHADOW_HALF_ROUND_TO_NEAREST == 1 249 v.ui += (v.ui & 0x3fff) != 0x1000 ? 0x1000 : 0;
251 v.ui -= expAdjust << fp32FractionBits;
252 }
else if (v.si <= infN) {
254 }
else if (v.si < nanN) {
259 return sign | (v.ui & 0x7fff);
263 MSHADOW_XINLINE uint16_t float2half(
const volatile float& value)
const volatile {
266 uint32_t sign = v.si & signN;
273 }
else if (v.si < minN) {
275 uint32_t exp32 = v.ui >> fp32FractionBits;
276 int32_t exp16 = exp32 - expAdjust;
279 uint32_t vshift = 1 - exp16;
280 uint32_t significand = fp32HiddenBit | (v.ui & fp32FractionMask);
281 v.ui = significand >> vshift;
282 #if MSHADOW_HALF_ROUND_TO_NEAREST == 1 284 v.ui += (v.ui & 0x3fff) != 0x1000 || (significand & 0x7ff) ? 0x1000 : 0;
286 }
else if (v.si <= maxN) {
288 #if MSHADOW_HALF_ROUND_TO_NEAREST == 1 290 v.ui += (v.ui & 0x3fff) != 0x1000 ? 0x1000 : 0;
292 v.ui -= expAdjust << fp32FractionBits;
293 }
else if (v.si <= infN) {
295 }
else if (v.si < nanN) {
300 return sign | (v.ui & 0x7fff);
306 int32_t sign = v.si & signC;
309 v.si ^= ((v.si + minD) ^ v.si) & -(v.si > subC);
310 v.si ^= ((v.si + maxD) ^ v.si) & -(v.si > maxC);
314 int32_t
mask = -(norC > v.si);
316 v.si ^= (s.si ^ v.si) & mask;
321 MSHADOW_XINLINE float half2float(
const volatile uint16_t& value)
const volatile {
324 int32_t sign = v.si & signC;
327 v.si ^= ((v.si + minD) ^ v.si) & -(v.si > subC);
328 v.si ^= ((v.si + maxD) ^ v.si) & -(v.si > maxC);
332 int32_t
mask = -(norC > v.si);
334 v.si ^= (s.si ^ v.si) & mask;
341 #if (MSHADOW_CUDA_HALF && defined(__CUDA_ARCH__)) 342 cuhalf_ = __float2half(
float(value));
343 #elif(MSHADOW_USE_F16C) 344 half_ = _cvtss_sh(static_cast<float>(value), 0);
346 half_ = float2half(
float(value));
368 #define MSHADOW_HALF_MIN mshadow::half::half_t::Binary(0xFBFF); 369 #define MSHADOW_HALF_MAX mshadow::half::half_t::Binary(0x7BFF); 370 #define MSHADOW_HALF_SIGN_BIT 0x8000 371 #define MSHADOW_HALF_EXPONENT_BITS 0x7c00 374 #endif // MSHADOW_HALF_H_ class MSHADOW_ALIGNED(2) half_t
Definition: half.h:112
#define MSHADOW_HALF_CONVERSIONOP(T)
Definition: half.h:95
#define MSHADOW_HALF_ASSIGNOP(AOP, OP)
Definition: half.h:76
MSHADOW_XINLINE half2_t operator+(half2_t a, half2_t b)
overloaded + operator for half2_t
Definition: half2.h:106
#define MSHADOW_XINLINE
Definition: base.h:230
MSHADOW_XINLINE half2_t operator-(half2_t a, half2_t b)
overloaded - operator for half2_t
Definition: half2.h:115
MaskExp< IndexExp, SrcExp, DType > mask(const Exp< IndexExp, DType, e1 > &index, const Exp< SrcExp, DType, e2 > &src)
Definition: mask.h:57
#define MSHADOW_HALF_OPERATOR(RTYPE, OP)
Definition: half.h:63
overloaded + operator between half_t and bf16_t
Definition: base.h:334