| 1 | //===-lib/fp_extend.h - low precision -> high precision conversion -*- C |
| 2 | //-*-===// |
| 3 | // |
| 4 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 5 | // See https://llvm.org/LICENSE.txt for license information. |
| 6 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // Set source and destination setting |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef FP_EXTEND_HEADER |
| 15 | #define |
| 16 | |
| 17 | #include "int_lib.h" |
| 18 | |
| 19 | #if defined SRC_SINGLE |
| 20 | typedef float src_t; |
| 21 | typedef uint32_t src_rep_t; |
| 22 | #define SRC_REP_C UINT32_C |
| 23 | static const int srcBits = sizeof(src_t) * CHAR_BIT; |
| 24 | static const int srcSigFracBits = 23; |
| 25 | // -1 accounts for the sign bit. |
| 26 | // srcBits - srcSigFracBits - 1 |
| 27 | static const int srcExpBits = 8; |
| 28 | #define src_rep_t_clz clzsi |
| 29 | |
| 30 | #elif defined SRC_DOUBLE |
| 31 | typedef double src_t; |
| 32 | typedef uint64_t src_rep_t; |
| 33 | #define SRC_REP_C UINT64_C |
| 34 | static const int srcBits = sizeof(src_t) * CHAR_BIT; |
| 35 | static const int srcSigFracBits = 52; |
| 36 | // -1 accounts for the sign bit. |
| 37 | // srcBits - srcSigFracBits - 1 |
| 38 | static const int srcExpBits = 11; |
| 39 | |
| 40 | static inline int src_rep_t_clz_impl(src_rep_t a) { return __builtin_clzll(a); } |
| 41 | #define src_rep_t_clz src_rep_t_clz_impl |
| 42 | |
| 43 | #elif defined SRC_80 |
| 44 | typedef xf_float src_t; |
| 45 | typedef __uint128_t src_rep_t; |
| 46 | #define SRC_REP_C (__uint128_t) |
| 47 | // sign bit, exponent and significand occupy the lower 80 bits. |
| 48 | static const int srcBits = 80; |
| 49 | static const int srcSigFracBits = 63; |
| 50 | // -1 accounts for the sign bit. |
| 51 | // -1 accounts for the explicitly stored integer bit. |
| 52 | // srcBits - srcSigFracBits - 1 - 1 |
| 53 | static const int srcExpBits = 15; |
| 54 | |
| 55 | #elif defined SRC_HALF |
| 56 | #ifdef COMPILER_RT_HAS_FLOAT16 |
| 57 | typedef _Float16 src_t; |
| 58 | #else |
| 59 | typedef uint16_t src_t; |
| 60 | #endif |
| 61 | typedef uint16_t src_rep_t; |
| 62 | #define SRC_REP_C UINT16_C |
| 63 | static const int srcBits = sizeof(src_t) * CHAR_BIT; |
| 64 | static const int srcSigFracBits = 10; |
| 65 | // -1 accounts for the sign bit. |
| 66 | // srcBits - srcSigFracBits - 1 |
| 67 | static const int srcExpBits = 5; |
| 68 | |
| 69 | static inline int src_rep_t_clz_impl(src_rep_t a) { |
| 70 | return __builtin_clz(a) - 16; |
| 71 | } |
| 72 | |
| 73 | #define src_rep_t_clz src_rep_t_clz_impl |
| 74 | |
| 75 | #elif defined SRC_BFLOAT16 |
| 76 | #ifdef COMPILER_RT_HAS_BFLOAT16 |
| 77 | typedef __bf16 src_t; |
| 78 | #else |
| 79 | typedef uint16_t src_t; |
| 80 | #endif |
| 81 | typedef uint16_t src_rep_t; |
| 82 | #define SRC_REP_C UINT16_C |
| 83 | static const int srcBits = sizeof(src_t) * CHAR_BIT; |
| 84 | static const int srcSigFracBits = 7; |
| 85 | // -1 accounts for the sign bit. |
| 86 | // srcBits - srcSigFracBits - 1 |
| 87 | static const int srcExpBits = 8; |
| 88 | #define src_rep_t_clz __builtin_clz |
| 89 | |
| 90 | #else |
| 91 | #error Source should be half, single, or double precision! |
| 92 | #endif // end source precision |
| 93 | |
| 94 | #if defined DST_SINGLE |
| 95 | typedef float dst_t; |
| 96 | typedef uint32_t dst_rep_t; |
| 97 | #define DST_REP_C UINT32_C |
| 98 | static const int dstBits = sizeof(dst_t) * CHAR_BIT; |
| 99 | static const int dstSigFracBits = 23; |
| 100 | // -1 accounts for the sign bit. |
| 101 | // dstBits - dstSigFracBits - 1 |
| 102 | static const int dstExpBits = 8; |
| 103 | |
| 104 | #elif defined DST_DOUBLE |
| 105 | typedef double dst_t; |
| 106 | typedef uint64_t dst_rep_t; |
| 107 | #define DST_REP_C UINT64_C |
| 108 | static const int dstBits = sizeof(dst_t) * CHAR_BIT; |
| 109 | static const int dstSigFracBits = 52; |
| 110 | // -1 accounts for the sign bit. |
| 111 | // dstBits - dstSigFracBits - 1 |
| 112 | static const int dstExpBits = 11; |
| 113 | |
| 114 | #elif defined DST_QUAD |
| 115 | typedef tf_float dst_t; |
| 116 | typedef __uint128_t dst_rep_t; |
| 117 | #define DST_REP_C (__uint128_t) |
| 118 | static const int dstBits = sizeof(dst_t) * CHAR_BIT; |
| 119 | static const int dstSigFracBits = 112; |
| 120 | // -1 accounts for the sign bit. |
| 121 | // dstBits - dstSigFracBits - 1 |
| 122 | static const int dstExpBits = 15; |
| 123 | |
| 124 | #else |
| 125 | #error Destination should be single, double, or quad precision! |
| 126 | #endif // end destination precision |
| 127 | |
| 128 | // End of specialization parameters. |
| 129 | |
| 130 | // TODO: These helper routines should be placed into fp_lib.h |
| 131 | // Currently they depend on macros/constants defined above. |
| 132 | |
| 133 | static inline src_rep_t (src_rep_t x) { |
| 134 | const src_rep_t srcSignMask = SRC_REP_C(1) << (srcBits - 1); |
| 135 | return (x & srcSignMask) >> (srcBits - 1); |
| 136 | } |
| 137 | |
| 138 | static inline src_rep_t (src_rep_t x) { |
| 139 | const int srcSigBits = srcBits - 1 - srcExpBits; |
| 140 | const src_rep_t srcExpMask = ((SRC_REP_C(1) << srcExpBits) - 1) << srcSigBits; |
| 141 | return (x & srcExpMask) >> srcSigBits; |
| 142 | } |
| 143 | |
| 144 | static inline src_rep_t (src_rep_t x) { |
| 145 | const src_rep_t srcSigFracMask = (SRC_REP_C(1) << srcSigFracBits) - 1; |
| 146 | return x & srcSigFracMask; |
| 147 | } |
| 148 | |
| 149 | #ifdef src_rep_t_clz |
| 150 | static inline int clz_in_sig_frac(src_rep_t sigFrac) { |
| 151 | const int skip = 1 + srcExpBits; |
| 152 | return src_rep_t_clz(sigFrac) - skip; |
| 153 | } |
| 154 | #endif |
| 155 | |
| 156 | static inline dst_rep_t construct_dst_rep(dst_rep_t sign, dst_rep_t exp, dst_rep_t sigFrac) { |
| 157 | return (sign << (dstBits - 1)) | (exp << (dstBits - 1 - dstExpBits)) | sigFrac; |
| 158 | } |
| 159 | |
| 160 | // Two helper routines for conversion to and from the representation of |
| 161 | // floating-point data as integer values follow. |
| 162 | |
| 163 | static inline src_rep_t srcToRep(src_t x) { |
| 164 | const union { |
| 165 | src_t f; |
| 166 | src_rep_t i; |
| 167 | } rep = {.f = x}; |
| 168 | return rep.i; |
| 169 | } |
| 170 | |
| 171 | static inline dst_t dstFromRep(dst_rep_t x) { |
| 172 | const union { |
| 173 | dst_t f; |
| 174 | dst_rep_t i; |
| 175 | } rep = {.i = x}; |
| 176 | return rep.f; |
| 177 | } |
| 178 | // End helper routines. Conversion implementation follows. |
| 179 | |
| 180 | #endif // FP_EXTEND_HEADER |
| 181 | |