• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 #include <stdint.h>
2 #include <float.h>
3 #include <math.h>
4 #include "atomic.h"
5 
6 #define ASUINT64(x) ((union {double f; uint64_t i;}){x}).i
7 #define ZEROINFNAN (0x7ff-0x3ff-52-1)
8 
9 struct num { uint64_t m; int e; int sign; };
10 
normalize(double x)11 static struct num normalize(double x)
12 {
13 	uint64_t ix = ASUINT64(x);
14 	int e = ix>>52;
15 	int sign = e & 0x800;
16 	e &= 0x7ff;
17 	if (!e) {
18 		ix = ASUINT64(x*0x1p63);
19 		e = ix>>52 & 0x7ff;
20 		e = e ? e-63 : 0x800;
21 	}
22 	ix &= (1ull<<52)-1;
23 	ix |= 1ull<<52;
24 	ix <<= 1;
25 	e -= 0x3ff + 52 + 1;
26 	return (struct num){ix,e,sign};
27 }
28 
mul(uint64_t * hi,uint64_t * lo,uint64_t x,uint64_t y)29 static void mul(uint64_t *hi, uint64_t *lo, uint64_t x, uint64_t y)
30 {
31 	uint64_t t1,t2,t3;
32 	uint64_t xlo = (uint32_t)x, xhi = x>>32;
33 	uint64_t ylo = (uint32_t)y, yhi = y>>32;
34 
35 	t1 = xlo*ylo;
36 	t2 = xlo*yhi + xhi*ylo;
37 	t3 = xhi*yhi;
38 	*lo = t1 + (t2<<32);
39 	*hi = t3 + (t2>>32) + (t1 > *lo);
40 }
41 
fma(double x,double y,double z)42 double fma(double x, double y, double z)
43 {
44 	#pragma STDC FENV_ACCESS ON
45 
46 	/* normalize so top 10bits and last bit are 0 */
47 	struct num nx, ny, nz;
48 	nx = normalize(x);
49 	ny = normalize(y);
50 	nz = normalize(z);
51 
52 	if (nx.e >= ZEROINFNAN || ny.e >= ZEROINFNAN)
53 		return x*y + z;
54 	if (nz.e >= ZEROINFNAN) {
55 		if (nz.e > ZEROINFNAN) /* z==0 */
56 			return x*y + z;
57 		return z;
58 	}
59 
60 	/* mul: r = x*y */
61 	uint64_t rhi, rlo, zhi, zlo;
62 	mul(&rhi, &rlo, nx.m, ny.m);
63 	/* either top 20 or 21 bits of rhi and last 2 bits of rlo are 0 */
64 
65 	/* align exponents */
66 	int e = nx.e + ny.e;
67 	int d = nz.e - e;
68 	/* shift bits z<<=kz, r>>=kr, so kz+kr == d, set e = e+kr (== ez-kz) */
69 	if (d > 0) {
70 		if (d < 64) {
71 			zlo = nz.m<<d;
72 			zhi = nz.m>>64-d;
73 		} else {
74 			zlo = 0;
75 			zhi = nz.m;
76 			e = nz.e - 64;
77 			d -= 64;
78 			if (d == 0) {
79 			} else if (d < 64) {
80 				rlo = rhi<<64-d | rlo>>d | !!(rlo<<64-d);
81 				rhi = rhi>>d;
82 			} else {
83 				rlo = 1;
84 				rhi = 0;
85 			}
86 		}
87 	} else {
88 		zhi = 0;
89 		d = -d;
90 		if (d == 0) {
91 			zlo = nz.m;
92 		} else if (d < 64) {
93 			zlo = nz.m>>d | !!(nz.m<<64-d);
94 		} else {
95 			zlo = 1;
96 		}
97 	}
98 
99 	/* add */
100 	int sign = nx.sign^ny.sign;
101 	int samesign = !(sign^nz.sign);
102 	int nonzero = 1;
103 	if (samesign) {
104 		/* r += z */
105 		rlo += zlo;
106 		rhi += zhi + (rlo < zlo);
107 	} else {
108 		/* r -= z */
109 		uint64_t t = rlo;
110 		rlo -= zlo;
111 		rhi = rhi - zhi - (t < rlo);
112 		if (rhi>>63) {
113 			rlo = -rlo;
114 			rhi = -rhi-!!rlo;
115 			sign = !sign;
116 		}
117 		nonzero = !!rhi;
118 	}
119 
120 	/* set rhi to top 63bit of the result (last bit is sticky) */
121 	if (nonzero) {
122 		e += 64;
123 		d = a_clz_64(rhi)-1;
124 		/* note: d > 0 */
125 		rhi = rhi<<d | rlo>>64-d | !!(rlo<<d);
126 	} else if (rlo) {
127 		d = a_clz_64(rlo)-1;
128 		if (d < 0)
129 			rhi = rlo>>1 | (rlo&1);
130 		else
131 			rhi = rlo<<d;
132 	} else {
133 		/* exact +-0 */
134 		return x*y + z;
135 	}
136 	e -= d;
137 
138 	/* convert to double */
139 	int64_t i = rhi; /* i is in [1<<62,(1<<63)-1] */
140 	if (sign)
141 		i = -i;
142 	double r = i; /* |r| is in [0x1p62,0x1p63] */
143 
144 	if (e < -1022-62) {
145 		/* result is subnormal before rounding */
146 		if (e == -1022-63) {
147 			double c = 0x1p63;
148 			if (sign)
149 				c = -c;
150 			if (r == c) {
151 				/* min normal after rounding, underflow depends
152 				   on arch behaviour which can be imitated by
153 				   a double to float conversion */
154 				float fltmin = 0x0.ffffff8p-63*FLT_MIN * r;
155 				return DBL_MIN/FLT_MIN * fltmin;
156 			}
157 			/* one bit is lost when scaled, add another top bit to
158 			   only round once at conversion if it is inexact */
159 			if (rhi << 53) {
160 				i = rhi>>1 | (rhi&1) | 1ull<<62;
161 				if (sign)
162 					i = -i;
163 				r = i;
164 				r = 2*r - c; /* remove top bit */
165 
166 				/* raise underflow portably, such that it
167 				   cannot be optimized away */
168 				{
169 					double_t tiny = DBL_MIN/FLT_MIN * r;
170 					r += (double)(tiny*tiny) * (r-r);
171 				}
172 			}
173 		} else {
174 			/* only round once when scaled */
175 			d = 10;
176 			i = ( rhi>>d | !!(rhi<<64-d) ) << d;
177 			if (sign)
178 				i = -i;
179 			r = i;
180 		}
181 	}
182 	return scalbn(r, e);
183 }
184