| 1 | #pragma once |
| 2 | |
| 3 | #include <bit> |
| 4 | #include <cstdint> |
| 5 | |
| 6 | inline float bf16_to_float(uint16_t bf16) { |
| 7 | uint32_t bits = static_cast<uint32_t>(bf16) << 16; |
| 8 | return std::bit_cast<float>(bits); |
| 9 | } |
| 10 | |
| 11 | inline uint16_t float_to_bf16(float fp32) { |
| 12 | uint32_t bits = std::bit_cast<uint32_t>(fp32); |
| 13 | return static_cast<uint16_t>(bits >> 16); |
| 14 | } |
| 15 | |
| 16 | /************************************************ |
| 17 | * Copyright (c) 2015 Princeton Vision Group |
| 18 | * Licensed under the MIT license. |
| 19 | * Codes below copied from |
| 20 | * https://github.com/PrincetonVision/marvin/tree/master/tools/tensorIO_matlab |
| 21 | *************************************************/ |
| 22 | inline uint16_t float_to_fp16(float fp32) { |
| 23 | uint16_t fp16; |
| 24 | |
| 25 | unsigned u, remainder, shift, lsb, lsb_s1, lsb_m1; |
| 26 | unsigned sign, exponent, mantissa; |
| 27 | |
| 28 | uint32_t x = std::bit_cast<uint32_t>(fp32); |
| 29 | u = (x & 0x7fffffff); |
| 30 | |
| 31 | // Get rid of +NaN/-NaN case first. |
| 32 | if (u > 0x7f800000) { |
| 33 | fp16 = 0x7fffU; |
| 34 | return fp16; |
| 35 | } |
| 36 | |
| 37 | sign = ((x >> 16) & 0x8000); |
| 38 | |
| 39 | // Get rid of +Inf/-Inf, +0/-0. |
| 40 | if (u > 0x477fefff) { |
| 41 | fp16 = sign | 0x7c00U; |
| 42 | return fp16; |
| 43 | } |
| 44 | if (u < 0x33000001) { |
| 45 | fp16 = (sign | 0x0000); |
| 46 | return fp16; |
| 47 | } |
| 48 | |
| 49 | exponent = ((u >> 23) & 0xff); |
| 50 | mantissa = (u & 0x7fffff); |
| 51 | |
| 52 | if (exponent > 0x70) { |
| 53 | shift = 13; |
| 54 | exponent -= 0x70; |
| 55 | } else { |
| 56 | shift = 0x7e - exponent; |
| 57 | exponent = 0; |
| 58 | mantissa |= 0x800000; |
| 59 | } |
| 60 | lsb = (1 << shift); |
| 61 | lsb_s1 = (lsb >> 1); |
| 62 | lsb_m1 = (lsb - 1); |
| 63 | |
| 64 | // Round to nearest even. |
| 65 | remainder = (mantissa & lsb_m1); |
| 66 | mantissa >>= shift; |
| 67 | if (remainder > lsb_s1 || (remainder == lsb_s1 && (mantissa & 0x1))) { |
| 68 | ++mantissa; |
| 69 | if (!(mantissa & 0x3ff)) { |
| 70 | ++exponent; |
| 71 | mantissa = 0; |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | fp16 = (sign | (exponent << 10) | mantissa); |
| 76 | |
| 77 | return fp16; |
| 78 | } |
| 79 | |
| 80 | inline float fp16_to_float(uint16_t fp16) { |
| 81 | unsigned sign = ((fp16 >> 15) & 1); |
| 82 | unsigned exponent = ((fp16 >> 10) & 0x1f); |
| 83 | unsigned mantissa = ((fp16 & 0x3ff) << 13); |
| 84 | uint32_t temp; |
| 85 | if (exponent == 0x1f) { /* NaN or Inf */ |
| 86 | mantissa = (mantissa ? (sign = 0, 0x7fffff) : 0); |
| 87 | exponent = 0xff; |
| 88 | } else if (!exponent) { /* Denorm or Zero */ |
| 89 | if (mantissa) { |
| 90 | unsigned int msb; |
| 91 | exponent = 0x71; |
| 92 | do { |
| 93 | msb = (mantissa & 0x400000); |
| 94 | mantissa <<= 1; /* normalize */ |
| 95 | --exponent; |
| 96 | } while (!msb); |
| 97 | mantissa &= 0x7fffff; /* 1.mantissa is implicit */ |
| 98 | } |
| 99 | } else { |
| 100 | exponent += 0x70; |
| 101 | } |
| 102 | temp = ((sign << 31) | (exponent << 23) | mantissa); |
| 103 | return std::bit_cast<float>(temp); |
| 104 | } |