cuda_half_t.h 9.81 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you 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.
 */

/*!
 * \file cuda_half_t.h
 * \brief half_t (fp16) definition for cuda codegen.
 */
#ifndef TVM_CODEGEN_LITERAL_CUDA_HALF_T_H_
#define TVM_CODEGEN_LITERAL_CUDA_HALF_T_H_

static constexpr const char* _cuda_half_t_def = R"(
typedef unsigned short uint16_t;
typedef unsigned char uint8_t;
30
typedef signed char int8_t;
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
typedef int int32_t;
typedef unsigned long long uint64_t;
typedef unsigned int uint32_t;

#define TVM_FORCE_INLINE inline __attribute__((always_inline))
#define TVM_XINLINE TVM_FORCE_INLINE __device__ __host__
#define TVM_ALIGNED(x) __attribute__ ((aligned(x)))
#define TVM_HALF_OPERATOR(RTYPE, OP)                              \
  TVM_XINLINE RTYPE operator OP (half a, half b) {                \
    return RTYPE(float(a) OP float(b));                           \
  }                                                               \
  template<typename T>                                            \
  TVM_XINLINE RTYPE operator OP (half a, T b) {                   \
    return RTYPE(float(a) OP float(b));                           \
  }                                                               \
  template<typename T>                                            \
  TVM_XINLINE RTYPE operator OP (T a, half b) {                   \
    return RTYPE(float(a) OP float(b));                           \
  }

#define TVM_HALF_ASSIGNOP(AOP, OP)                                \
  template<typename T>                                            \
  TVM_XINLINE half operator AOP (const T& a) {                    \
    return *this = half(float(*this) OP float(a));                \
  }                                                               \
  template<typename T>                                            \
  TVM_XINLINE half operator AOP (const volatile T& a) volatile {  \
    return *this = half(float(*this) OP float(a));                \
  }

class TVM_ALIGNED(2) half {
 public:
  uint16_t half_;

  static TVM_XINLINE half Binary(uint16_t value) {
    half res;
    res.half_ = value;
    return res;
  }

  TVM_XINLINE half() {}

  TVM_XINLINE half(const float& value) { constructor(value); }
  TVM_XINLINE explicit half(const double& value) { constructor(value); }
  TVM_XINLINE explicit half(const int8_t& value) { constructor(value); }
  TVM_XINLINE explicit half(const uint8_t& value) { constructor(value); }
  TVM_XINLINE explicit half(const int32_t& value) { constructor(value); }
  TVM_XINLINE explicit half(const uint32_t& value) { constructor(value); }
79
  TVM_XINLINE explicit half(const long long& value) { constructor(value); }
80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
  TVM_XINLINE explicit half(const uint64_t& value) { constructor(value); }

  TVM_XINLINE operator float() const {                          \
    return float(half2float(half_));                            \
  }                                                             \
  TVM_XINLINE operator float() const volatile {                 \
    return float(half2float(half_));                            \
  }


  TVM_HALF_ASSIGNOP(+=, +)
  TVM_HALF_ASSIGNOP(-=, -)
  TVM_HALF_ASSIGNOP(*=, *)
  TVM_HALF_ASSIGNOP(/=, /)

  TVM_XINLINE half operator+() {
    return *this;
  }

  TVM_XINLINE half operator-() {
    return half(-float(*this));
  }

  TVM_XINLINE half operator=(const half& a) {
    half_ = a.half_;
    return a;
  }

  template<typename T>
  TVM_XINLINE half operator=(const T& a) {
    return *this = half(a);
  }

  TVM_XINLINE half operator=(const half& a) volatile {
    half_ = a.half_;
    return a;
  }

  template<typename T>
  TVM_XINLINE half operator=(const T& a) volatile {
    return *this = half(a);
  }

 private:
  union Bits {
    float f;
    int32_t si;
    uint32_t ui;
  };

  static int const fp16FractionBits = 10;
  static int const fp32FractionBits = 23;
  static int32_t const fp32FractionMask = ~(~0u << fp32FractionBits);   // == 0x7fffff
  static int32_t const fp32HiddenBit = 1 << fp32FractionBits;   // == 0x800000
  static int const shift = fp32FractionBits - fp16FractionBits;   // == 13
  static int const shiftSign = 16;
  static int32_t const expAdjust = 127 - 15;   // exp32-127 = exp16-15, so exp16 = exp32 - (127-15)

  static int32_t const infN = 0x7F800000;   // flt32 infinity
  static int32_t const maxN = 0x477FFFFF;   // max flt32 that's a flt16 normal after >> by shift
  static int32_t const minN = 0x38800000;   // min flt16 normal as a flt32
  static int32_t const maxZ = 0x33000000;   // max fp32 number that's still rounded to zero in fp16
  static int32_t const signN = 0x80000000;  // flt32 sign bit

  static int32_t const infC = infN >> shift;
  static int32_t const nanN = (infC + 1) << shift;   // minimum flt16 nan as a flt32
  static int32_t const maxC = maxN >> shift;
  static int32_t const minC = minN >> shift;
  static int32_t const signC = signN >> shiftSign;  // flt16 sign bit

  static int32_t const mulN = 0x52000000;  // (1 << 23) / minN
  static int32_t const mulC = 0x33800000;  // minN / (1 << (23 - shift))

  static int32_t const subC = 0x003FF;  // max flt32 subnormal down shifted
  static int32_t const norC = 0x00400;  // min flt32 normal down shifted

  static int32_t const maxD = infC - maxC - 1;
  static int32_t const minD = minC - subC - 1;

  TVM_XINLINE uint16_t float2half(const float& value) const {
    Bits v;
    v.f = value;
    uint32_t sign = v.si & signN;    // grab sign bit
    v.si ^= sign;                    // clear sign bit from v
    sign >>= shiftSign;              // logical shift sign to fp16 position

    if (v.si <= maxZ) {
      // Handle eventual zeros here to ensure
      // vshift will not exceed 32 below.
      v.ui = 0;
    } else if (v.si < minN) {
      // Handle denorms
      uint32_t exp32 = v.ui >> fp32FractionBits;
      int32_t exp16 = exp32 - expAdjust;
      // If exp16 == 0 (just into the denorm range), then significant should be shifted right 1.
      // Smaller (so negative) exp16 values should result in greater right shifts.
      uint32_t vshift = 1 - exp16;
      uint32_t significand = fp32HiddenBit | (v.ui & fp32FractionMask);
      v.ui = significand >> vshift;
179
      v.ui += (v.ui & 0x3fff) != 0x1000 || (significand & 0x7ff) ? 0x1000 : 0;
180 181
    } else if (v.si <= maxN) {
      // Handle norms
182
      v.ui += (v.ui & 0x3fff) != 0x1000 ? 0x1000 : 0;
183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
      v.ui -= expAdjust << fp32FractionBits;
    } else if (v.si <= infN) {
      v.si = infN;
    } else if (v.si < nanN) {
      v.si = nanN;
    }

    v.ui >>= shift;
    return sign | (v.ui & 0x7fff);
  }

  // Same as above routine, except for addition of volatile keyword
  TVM_XINLINE uint16_t float2half(
    const volatile float& value) const volatile {
    Bits v;
    v.f = value;
    uint32_t sign = v.si & signN;    // grab sign bit
    v.si ^= sign;                    // clear sign bit from v
    sign >>= shiftSign;              // logical shift sign to fp16 position

    if (v.si <= maxZ) {
      // Handle eventual zeros here to ensure
      // vshift will not exceed 32 below.
      v.ui = 0;
    } else if (v.si < minN) {
      // Handle denorms
      uint32_t exp32 = v.ui >> fp32FractionBits;
      int32_t exp16 = exp32 - expAdjust;
      // If exp16 == 0 (just into the denorm range), then significant should be shifted right 1.
      // Smaller (so negative) exp16 values should result in greater right shifts.
      uint32_t vshift = 1 - exp16;
      uint32_t significand = fp32HiddenBit | (v.ui & fp32FractionMask);
      v.ui = significand >> vshift;
216
      v.ui += (v.ui & 0x3fff) != 0x1000 || (significand & 0x7ff) ? 0x1000 : 0;
217 218
    } else if (v.si <= maxN) {
      // Handle norms
219
      v.ui += (v.ui & 0x3fff) != 0x1000 ? 0x1000 : 0;
220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281
      v.ui -= expAdjust << fp32FractionBits;
    } else if (v.si <= infN) {
      v.si = infN;
    } else if (v.si < nanN) {
      v.si = nanN;
    }

    v.ui >>= shift;
    return sign | (v.ui & 0x7fff);
  }

  TVM_XINLINE float half2float(const uint16_t& value) const {
    Bits v;
    v.ui = value;
    int32_t sign = v.si & signC;
    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;
  }

  TVM_XINLINE float half2float(
    const volatile uint16_t& value) const volatile {
    Bits v;
    v.ui = value;
    int32_t sign = v.si & signC;
    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;
  }

  template<typename T>
  TVM_XINLINE void constructor(const T& value) {
    half_ = float2half(float(value));
  }
};

TVM_HALF_OPERATOR(half, +)
TVM_HALF_OPERATOR(half, -)
TVM_HALF_OPERATOR(half, *)
TVM_HALF_OPERATOR(half, /)
TVM_HALF_OPERATOR(bool, >)
TVM_HALF_OPERATOR(bool, <)
TVM_HALF_OPERATOR(bool, >=)
TVM_HALF_OPERATOR(bool, <=)
282 283 284 285

TVM_XINLINE half __float2half_rn(const float a) {
  return half(a);
}
286 287 288
)";

#endif  // TVM_CODEGEN_LITERAL_CUDA_HALF_T_H_