1 //=== lib/fp_trunc.h - high precision -> low precision conversion *- C -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Set source and destination precision setting
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #ifndef FP_TRUNC_HEADER
14 #define FP_TRUNC_HEADER
15 
16 #include "int_lib.h"
17 
18 #if defined SRC_SINGLE
19 typedef float src_t;
20 typedef uint32_t src_rep_t;
21 #define SRC_REP_C UINT32_C
22 static const int srcBits = sizeof(src_t) * CHAR_BIT;
23 static const int srcSigFracBits = 23;
24 // -1 accounts for the sign bit.
25 // srcBits - srcSigFracBits - 1
26 static const int srcExpBits = 8;
27 
28 #elif defined SRC_DOUBLE
29 typedef double src_t;
30 typedef uint64_t src_rep_t;
31 #define SRC_REP_C UINT64_C
32 static const int srcBits = sizeof(src_t) * CHAR_BIT;
33 static const int srcSigFracBits = 52;
34 // -1 accounts for the sign bit.
35 // srcBits - srcSigFracBits - 1
36 static const int srcExpBits = 11;
37 
38 #elif defined SRC_QUAD
39 typedef tf_float src_t;
40 typedef __uint128_t src_rep_t;
41 #define SRC_REP_C (__uint128_t)
42 static const int srcBits = sizeof(src_t) * CHAR_BIT;
43 static const int srcSigFracBits = 112;
44 // -1 accounts for the sign bit.
45 // srcBits - srcSigFracBits - 1
46 static const int srcExpBits = 15;
47 
48 #else
49 #error Source should be double precision or quad precision!
50 #endif // end source precision
51 
52 #if defined DST_DOUBLE
53 typedef double dst_t;
54 typedef uint64_t dst_rep_t;
55 #define DST_REP_C UINT64_C
56 static const int dstBits = sizeof(dst_t) * CHAR_BIT;
57 static const int dstSigFracBits = 52;
58 // -1 accounts for the sign bit.
59 // dstBits - dstSigFracBits - 1
60 static const int dstExpBits = 11;
61 
62 #elif defined DST_80
63 typedef xf_float dst_t;
64 typedef __uint128_t dst_rep_t;
65 #define DST_REP_C (__uint128_t)
66 static const int dstBits = 80;
67 static const int dstSigFracBits = 63;
68 // -1 accounts for the sign bit.
69 // -1 accounts for the explicitly stored integer bit.
70 // dstBits - dstSigFracBits - 1 - 1
71 static const int dstExpBits = 15;
72 
73 #elif defined DST_SINGLE
74 typedef float dst_t;
75 typedef uint32_t dst_rep_t;
76 #define DST_REP_C UINT32_C
77 static const int dstBits = sizeof(dst_t) * CHAR_BIT;
78 static const int dstSigFracBits = 23;
79 // -1 accounts for the sign bit.
80 // dstBits - dstSigFracBits - 1
81 static const int dstExpBits = 8;
82 
83 #elif defined DST_HALF
84 #ifdef COMPILER_RT_HAS_FLOAT16
85 typedef _Float16 dst_t;
86 #else
87 typedef uint16_t dst_t;
88 #endif
89 typedef uint16_t dst_rep_t;
90 #define DST_REP_C UINT16_C
91 static const int dstBits = sizeof(dst_t) * CHAR_BIT;
92 static const int dstSigFracBits = 10;
93 // -1 accounts for the sign bit.
94 // dstBits - dstSigFracBits - 1
95 static const int dstExpBits = 5;
96 
97 #elif defined DST_BFLOAT
98 typedef __bf16 dst_t;
99 typedef uint16_t dst_rep_t;
100 #define DST_REP_C UINT16_C
101 static const int dstBits = sizeof(dst_t) * CHAR_BIT;
102 static const int dstSigFracBits = 7;
103 // -1 accounts for the sign bit.
104 // dstBits - dstSigFracBits - 1
105 static const int dstExpBits = 8;
106 
107 #else
108 #error Destination should be single precision or double precision!
109 #endif // end destination precision
110 
111 // TODO: These helper routines should be placed into fp_lib.h
112 // Currently they depend on macros/constants defined above.
113 
extract_sign_from_src(src_rep_t x)114 static inline src_rep_t extract_sign_from_src(src_rep_t x) {
115   const src_rep_t srcSignMask = SRC_REP_C(1) << (srcBits - 1);
116   return (x & srcSignMask) >> (srcBits - 1);
117 }
118 
extract_exp_from_src(src_rep_t x)119 static inline src_rep_t extract_exp_from_src(src_rep_t x) {
120   const int srcSigBits = srcBits - 1 - srcExpBits;
121   const src_rep_t srcExpMask = ((SRC_REP_C(1) << srcExpBits) - 1) << srcSigBits;
122   return (x & srcExpMask) >> srcSigBits;
123 }
124 
extract_sig_frac_from_src(src_rep_t x)125 static inline src_rep_t extract_sig_frac_from_src(src_rep_t x) {
126   const src_rep_t srcSigFracMask = (SRC_REP_C(1) << srcSigFracBits) - 1;
127   return x & srcSigFracMask;
128 }
129 
construct_dst_rep(dst_rep_t sign,dst_rep_t exp,dst_rep_t sigFrac)130 static inline dst_rep_t construct_dst_rep(dst_rep_t sign, dst_rep_t exp, dst_rep_t sigFrac) {
131   dst_rep_t result = (sign << (dstBits - 1)) | (exp << (dstBits - 1 - dstExpBits)) | sigFrac;
132   // Set the explicit integer bit in F80 if present.
133   if (dstBits == 80 && exp) {
134     result |= (DST_REP_C(1) << dstSigFracBits);
135   }
136   return result;
137 }
138 
139 // End of specialization parameters.  Two helper routines for conversion to and
140 // from the representation of floating-point data as integer values follow.
141 
srcToRep(src_t x)142 static inline src_rep_t srcToRep(src_t x) {
143   const union {
144     src_t f;
145     src_rep_t i;
146   } rep = {.f = x};
147   return rep.i;
148 }
149 
dstFromRep(dst_rep_t x)150 static inline dst_t dstFromRep(dst_rep_t x) {
151   const union {
152     dst_t f;
153     dst_rep_t i;
154   } rep = {.i = x};
155   return rep.f;
156 }
157 
158 #endif // FP_TRUNC_HEADER
159