8 #ifndef MSHADOW_HALF_H_ 9 #define MSHADOW_HALF_H_ 13 #include <x86intrin.h> 14 #endif // MSHADOW_USE_F16C 18 #ifndef MSHADOW_HALF_ROUND_TO_NEAREST 19 #define MSHADOW_HALF_ROUND_TO_NEAREST 1 22 #if (MSHADOW_USE_CUDA && CUDA_VERSION >= 7050) 23 #define MSHADOW_CUDA_HALF 1 24 #include <cuda_fp16.h> 25 #if defined(__CUDA_ARCH__) 27 __host__ __device__
float __half2float_warp(
const volatile __half& h) {
29 #if CUDA_VERSION >= 9000 30 val =
const_cast<__half&
>(h);
34 return __half2float(val);
38 #define MSHADOW_CUDA_HALF 0 45 #define MSHADOW_HALF_OPERATOR(RTYPE, OP) \ 46 MSHADOW_XINLINE RTYPE operator OP (half_t a, half_t b) { \ 47 return RTYPE(float(a) OP float(b)); \ 49 template<typename T> \ 50 MSHADOW_XINLINE RTYPE operator OP (half_t a, T b) { \ 51 return RTYPE(float(a) OP float(b)); \ 53 template<typename T> \ 54 MSHADOW_XINLINE RTYPE operator OP (T a, half_t b) { \ 55 return RTYPE(float(a) OP float(b)); \ 58 #define MSHADOW_HALF_ASSIGNOP(AOP, OP) \ 59 template<typename T> \ 60 MSHADOW_XINLINE half_t operator AOP (const T& a) { \ 61 return *this = half_t(float(*this) OP float(a)); \ 63 template<typename T> \ 64 MSHADOW_XINLINE half_t operator AOP (const volatile T& a) volatile { \ 65 return *this = half_t(float(*this) OP float(a)); \ 68 #if (MSHADOW_CUDA_HALF && defined(__CUDA_ARCH__)) 69 #define MSHADOW_HALF_CONVERSIONOP(T) \ 70 MSHADOW_XINLINE operator T() const { \ 71 return T(__half2float(cuhalf_)); \ 73 MSHADOW_XINLINE operator T() const volatile { \ 74 return T(__half2float_warp(cuhalf_)); \ 76 #elif(MSHADOW_USE_F16C) 77 #define MSHADOW_HALF_CONVERSIONOP(T) \ 78 MSHADOW_XINLINE operator T() const { \ 79 return T(_cvtsh_ss(half_)); \ 81 MSHADOW_XINLINE operator T() const volatile { \ 82 return T(_cvtsh_ss(half_)); \ 85 #define MSHADOW_HALF_CONVERSIONOP(T) \ 86 MSHADOW_XINLINE operator T() const { \ 87 return T(half2float(half_)); \ 89 MSHADOW_XINLINE operator T() const volatile { \ 90 return T(half2float(half_)); \ 92 #endif // (MSHADOW_CUDA_HALF && defined(__CUDA_ARCH__)) 100 #endif // MSHADOW_CUDA_HALF 111 #if MSHADOW_CUDA_HALF 115 #endif // MSHADOW_CUDA_HALF 118 MSHADOW_XINLINE explicit half_t(
const double& value) { constructor(value); }
119 MSHADOW_XINLINE explicit half_t(
const int8_t& value) { constructor(value); }
120 MSHADOW_XINLINE explicit half_t(
const uint8_t& value) { constructor(value); }
121 MSHADOW_XINLINE explicit half_t(
const int32_t& value) { constructor(value); }
122 MSHADOW_XINLINE explicit half_t(
const uint32_t& value) { constructor(value); }
123 MSHADOW_XINLINE explicit half_t(
const int64_t& value) { constructor(value); }
124 MSHADOW_XINLINE explicit half_t(
const uint64_t& value) { constructor(value); }
138 return half_t(-
float(*
this));
148 return *
this = half_t(a);
158 return *
this = half_t(a);
168 static int const fp16FractionBits = 10;
169 static int const fp32FractionBits = 23;
170 static int32_t
const fp32FractionMask = ~(~0u << fp32FractionBits);
171 static int32_t
const fp32HiddenBit = 1 << fp32FractionBits;
172 static int const shift = fp32FractionBits - fp16FractionBits;
173 static int const shiftSign = 16;
174 static int32_t
const expAdjust = 127 - 15;
176 static int32_t
const infN = 0x7F800000;
177 static int32_t
const maxN = 0x477FFFFF;
178 static int32_t
const minN = 0x38800000;
179 static int32_t
const maxZ = 0x33000000;
180 static int32_t
const signN = 0x80000000;
182 static int32_t
const infC = infN >> shift;
183 static int32_t
const nanN = (infC + 1) << shift;
184 static int32_t
const maxC = maxN >> shift;
185 static int32_t
const minC = minN >> shift;
186 static int32_t
const signC = signN >> shiftSign;
188 static int32_t
const mulN = 0x52000000;
189 static int32_t
const mulC = 0x33800000;
191 static int32_t
const subC = 0x003FF;
192 static int32_t
const norC = 0x00400;
194 static int32_t
const maxD = infC - maxC - 1;
195 static int32_t
const minD = minC - subC - 1;
200 uint32_t sign = v.si & signN;
207 }
else if (v.si < minN) {
209 uint32_t exp32 = v.ui >> fp32FractionBits;
210 int32_t exp16 = exp32 - expAdjust;
213 uint32_t vshift = 1 - exp16;
214 uint32_t significand = fp32HiddenBit | (v.ui & fp32FractionMask);
215 v.ui = significand >> vshift;
223 #if MSHADOW_HALF_ROUND_TO_NEAREST == 1 225 v.ui += (v.ui & 0x3fff) != 0x1000 || (significand & 0x7ff) ? 0x1000 : 0;
227 }
else if (v.si <= maxN) {
229 #if MSHADOW_HALF_ROUND_TO_NEAREST == 1 231 v.ui += (v.ui & 0x3fff) != 0x1000 ? 0x1000 : 0;
233 v.ui -= expAdjust << fp32FractionBits;
234 }
else if (v.si <= infN) {
236 }
else if (v.si < nanN) {
241 return sign | (v.ui & 0x7fff);
245 MSHADOW_XINLINE uint16_t float2half(
const volatile float& value)
const volatile {
248 uint32_t sign = v.si & signN;
255 }
else if (v.si < minN) {
257 uint32_t exp32 = v.ui >> fp32FractionBits;
258 int32_t exp16 = exp32 - expAdjust;
261 uint32_t vshift = 1 - exp16;
262 uint32_t significand = fp32HiddenBit | (v.ui & fp32FractionMask);
263 v.ui = significand >> vshift;
264 #if MSHADOW_HALF_ROUND_TO_NEAREST == 1 266 v.ui += (v.ui & 0x3fff) != 0x1000 || (significand & 0x7ff) ? 0x1000 : 0;
268 }
else if (v.si <= maxN) {
270 #if MSHADOW_HALF_ROUND_TO_NEAREST == 1 272 v.ui += (v.ui & 0x3fff) != 0x1000 ? 0x1000 : 0;
274 v.ui -= expAdjust << fp32FractionBits;
275 }
else if (v.si <= infN) {
277 }
else if (v.si < nanN) {
282 return sign | (v.ui & 0x7fff);
288 int32_t sign = v.si & signC;
291 v.si ^= ((v.si + minD) ^ v.si) & -(v.si > subC);
292 v.si ^= ((v.si + maxD) ^ v.si) & -(v.si > maxC);
296 int32_t
mask = -(norC > v.si);
298 v.si ^= (s.si ^ v.si) & mask;
303 MSHADOW_XINLINE float half2float(
const volatile uint16_t& value)
const volatile {
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;
323 #if (MSHADOW_CUDA_HALF && defined(__CUDA_ARCH__)) 324 cuhalf_ = __float2half(
float(value));
325 #elif(MSHADOW_USE_F16C) 326 half_ = _cvtss_sh(static_cast<float>(value), 0);
328 half_ = float2half(
float(value));
350 #define MSHADOW_HALF_MIN mshadow::half::half_t::Binary(0xFBFF); 351 #define MSHADOW_HALF_MAX mshadow::half::half_t::Binary(0x7BFF); 352 #define MSHADOW_HALF_SIGN_BIT 0x8000 353 #define MSHADOW_HALF_EXPONENT_BITS 0x7c00 356 #endif // MSHADOW_HALF_H_ class MSHADOW_ALIGNED(2) half_t
Definition: half.h:94
#define MSHADOW_HALF_CONVERSIONOP(T)
Definition: half.h:77
#define MSHADOW_HALF_ASSIGNOP(AOP, OP)
Definition: half.h:58
MSHADOW_XINLINE half2_t operator+(half2_t a, half2_t b)
overloaded + operator for half2_t
Definition: half2.h:88
#define MSHADOW_XINLINE
Definition: base.h:204
MSHADOW_XINLINE half2_t operator-(half2_t a, half2_t b)
overloaded - operator for half2_t
Definition: half2.h:97
MaskExp< IndexExp, SrcExp, DType > mask(const Exp< IndexExp, DType, e1 > &index, const Exp< SrcExp, DType, e2 > &src)
Definition: mask.h:39
#define MSHADOW_HALF_OPERATOR(RTYPE, OP)
Definition: half.h:45
namespace for mshadow
Definition: base.h:282