xref: /aosp_15_r20/external/pytorch/c10/util/Float8_e5m2.h (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1 #pragma once
2 
3 /// Defines the Float8_e5m2 type (8-bit floating-point) including conversions
4 /// to standard C types and basic arithmetic operations. Note that arithmetic
5 /// operations are implemented by converting to floating point and
6 /// performing the operation in float32.
7 /// Binary configuration:
8 /// s eeeee mm
9 /// 1 sign bit
10 /// 5 exponent bits
11 /// 2 mantissa bits
12 /// bias = 15
13 ///
14 /// Implementation based on the paper https://arxiv.org/pdf/2209.05433.pdf
15 /// and inspired by Half implementation from pytorch/c10/util/Half.h
16 
17 #include <c10/util/Half.h>
18 
19 namespace c10 {
20 
21 namespace detail {
22 
23 /*
24  * Convert a 8-bit floating-point number in fp8 E5M2 format, in bit
25  * representation, to a 32-bit floating-point number in IEEE single-precision
26  * format, in bit representation.
27  *
28  * @note The implementation doesn't use any floating-point operations.
29  */
fp8e5m2_to_fp32_value(uint8_t input)30 inline C10_HOST_DEVICE float fp8e5m2_to_fp32_value(uint8_t input) {
31   /*
32    * Extend the fp8 E5M2 number to 32 bits and shift to the
33    * upper part of the 32-bit word:
34    *      +---+----+---+-----------------------------+
35    *      | S |EEEEE|MM|0000 0000 0000 0000 0000 0000|
36    *      +---+----+---+-----------------------------+
37    * Bits  31 26-30 24-25          0-23
38    *
39    * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0
40    * - zero bits.
41    */
42   uint16_t half_representation = input;
43   half_representation <<= 8;
44   return fp16_ieee_to_fp32_value(half_representation);
45 }
46 
47 /*
48  * Convert a 32-bit floating-point number in IEEE single-precision format to a
49  * 8-bit floating-point number in fp8 E5M2 format, in bit representation.
50  */
fp8e5m2_from_fp32_value(float f)51 inline C10_HOST_DEVICE uint8_t fp8e5m2_from_fp32_value(float f) {
52   /*
53    * Binary representation of fp32 infinity
54    * 0 11111111 00000000000000000000000
55    */
56   constexpr uint32_t fp32_inf = UINT32_C(255) << 23;
57 
58   /*
59    * Binary representation of 65536.0f, which is the first value
60    * not representable in fp8e5m2 range:
61    * 0 11111 00 - fp8e5m2
62    * 0 10001111 00000000000000000000000 - fp32
63    */
64   constexpr uint32_t fp8_max = UINT32_C(143) << 23;
65 
66   /*
67    * A mask for converting fp32 numbers lower than fp8e5m2 normal range
68    * into denorm representation
69    * magic number: ((127 - 15) + (23 - 2) + 1)
70    */
71   constexpr uint32_t denorm_mask = UINT32_C(134) << 23;
72 
73   uint32_t f_bits = fp32_to_bits(f);
74   uint8_t result = 0u;
75 
76   /*
77    * Extract the sign of the input number into the high bit of the 32-bit word:
78    *
79    *      +---+----------------------------------+
80    *      | S |0000000 00000000 00000000 00000000|
81    *      +---+----------------------------------+
82    * Bits  31                 0-31
83    */
84   const uint32_t sign = f_bits & UINT32_C(0x80000000);
85 
86   /*
87    * Set sign bit to 0
88    */
89   f_bits ^= sign;
90 
91   if (f_bits >= fp8_max) {
92     // NaN - all exponent and mantissa bits set to 1
93     result = f_bits > fp32_inf ? UINT8_C(0x7F) : UINT8_C(0x7C);
94   } else {
95     if (f_bits < (UINT32_C(113) << 23)) {
96       // Input number is smaller than 2^(-14), which is the smallest
97       // fp8e5m2 normal number
98       f_bits =
99           fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask));
100       result = static_cast<uint8_t>(f_bits - denorm_mask);
101     } else {
102       // resulting mantissa is odd
103       uint32_t mant_odd = (f_bits >> 21) & 1;
104 
105       // update exponent, rounding bias part 1
106       f_bits += ((uint32_t)(15 - 127) << 23) + 0xFFFFF;
107 
108       // rounding bias part 2
109       f_bits += mant_odd;
110 
111       // take the bits!
112       result = static_cast<uint8_t>(f_bits >> 21);
113     }
114   }
115 
116   result |= static_cast<uint8_t>(sign >> 24);
117   return result;
118 }
119 
120 } // namespace detail
121 
122 struct alignas(1) Float8_e5m2 {
123   uint8_t x;
124 
125   struct from_bits_t {};
from_bitsFloat8_e5m2126   C10_HOST_DEVICE static constexpr from_bits_t from_bits() {
127     return from_bits_t();
128   }
129 
130   Float8_e5m2() = default;
131 
Float8_e5m2Float8_e5m2132   constexpr C10_HOST_DEVICE Float8_e5m2(uint8_t bits, from_bits_t) : x(bits) {}
133   inline C10_HOST_DEVICE Float8_e5m2(float value);
134   inline C10_HOST_DEVICE operator float() const;
135   inline C10_HOST_DEVICE bool isnan() const;
136   inline C10_HOST_DEVICE bool isinf() const;
137 };
138 
139 C10_API inline std::ostream& operator<<(
140     std::ostream& out,
141     const Float8_e5m2& value) {
142   out << (float)value;
143   return out;
144 }
145 
146 } // namespace c10
147 
148 #include <c10/util/Float8_e5m2-inl.h> // IWYU pragma: keep
149