4
* Copyright (c) 1998, 2001, Oracle and/or its affiliates. All rights reserved.
5
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7
* This code is free software; you can redistribute it and/or modify it
8
* under the terms of the GNU General Public License version 2 only, as
9
* published by the Free Software Foundation. Oracle designates this
10
* particular file as subject to the "Classpath" exception as provided
11
* by Oracle in the LICENSE file that accompanied this code.
13
* This code is distributed in the hope that it will be useful, but WITHOUT
14
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
15
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16
* version 2 for more details (a copy is included in the LICENSE file that
17
* accompanied this code).
19
* You should have received a copy of the GNU General Public License version
20
* 2 along with this work; if not, write to the Free Software Foundation,
21
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
23
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
24
* or visit www.oracle.com if you need additional information or have any
29
* __kernel_rem_pio2(x,y,e0,nx,prec,ipio2)
30
* double x[],y[]; int e0,nx,prec; int ipio2[];
32
* __kernel_rem_pio2 return the last three digits of N with
36
* The method is to compute the integer (mod 8) and fraction parts of
37
* (2/pi)*x without doing the full multiplication. In general we
38
* skip the part of the product that are known to be a huge integer (
39
* more accurately, = 0 mod 8 ). Thus the number of operations are
40
* independent of the exponent of the input.
42
* (2/pi) is represented by an array of 24-bit integers in ipio2[].
45
* x[] The input value (must be positive) is broken into nx
46
* pieces of 24-bit integers in double precision format.
47
* x[i] will be the i-th 24 bit of x. The scaled exponent
48
* of x[0] is given in input parameter e0 (i.e., x[0]*2^e0
49
* match x's up to 24 bits.
51
* Example of breaking a double positive z into x[0]+x[1]+x[2]:
59
* y[] ouput result in an array of double precision numbers.
60
* The dimension of y[] is:
65
* The actual value is the sum of them. Thus for 113-bit
66
* precison, one may have to do something like:
68
* long double t,w,r_head, r_tail;
69
* t = (long double)y[2] + (long double)y_1_;
70
* w = (long double)y_0_;
72
* r_tail = w - (r_head - t);
74
* e0 The exponent of x[0]
78
* prec an integer indicating the precision:
81
* 2 64 bits (extended)
85
* integer array, contains the (24*i)-th to (24*i+23)-th
86
* bit of 2/pi after binary point. The corresponding
89
* ipio2[i] * 2^(-24(i+1)).
92
* double scalbn(), floor();
95
* Here is the description of some local variables:
97
* jk jk+1 is the initial number of terms of ipio2[] needed
98
* in the computation. The recommended value is 2,3,4,
99
* 6 for single, double, extended,and quad.
101
* jz local integer variable indicating the number of
102
* terms of ipio2[] used.
106
* jv index for pointing to the suitable ipio2[] for the
107
* computation. In general, we want
108
* ( 2^e0*x[0] * ipio2[jv-1]*2^(-24jv) )/8
109
* is an integer. Thus
110
* e0-3-24*jv >= 0 or (e0-3)/24 >= jv
111
* Hence jv = max(0,(e0-3)/24).
113
* jp jp+1 is the number of terms in PIo2[] needed, jp = jk.
115
* q[] double array with integral value, representing the
116
* 24-bits chunk of the product of x and 2/pi.
118
* q0 the corresponding exponent of q[0]. Note that the
119
* exponent for q[i] would be q0-24*i.
121
* PIo2[] double precision array, obtained by cutting pi/2
122
* into 24 bits chunks.
124
* f[] ipio2[] in floating point
126
* iq[] integer array by breaking up q[] in 24-bits chunk.
128
* fq[] final product of x*(2/pi) in fq[0],..,fq[jk]
130
* ih integer. If >0 it indicates q[] is >= 0.5, hence
131
* it also indicates the *sign* of the result.
138
* The hexadecimal values are the intended ones for the following
139
* constants. The decimal values may be used, provided that the
140
* compiler will convert from decimal to binary accurately enough
141
* to produce the hexadecimal values shown.
144
static partial class fdlibm
146
static readonly int[] init_jk = {2,3,4,6}; /* initial value for jk */
148
static readonly double[] PIo2 = {
149
1.57079625129699707031e+00, /* 0x3FF921FB, 0x40000000 */
150
7.54978941586159635335e-08, /* 0x3E74442D, 0x00000000 */
151
5.39030252995776476554e-15, /* 0x3CF84698, 0x80000000 */
152
3.28200341580791294123e-22, /* 0x3B78CC51, 0x60000000 */
153
1.27065575308067607349e-29, /* 0x39F01B83, 0x80000000 */
154
1.22933308981111328932e-36, /* 0x387A2520, 0x40000000 */
155
2.73370053816464559624e-44, /* 0x36E38222, 0x80000000 */
156
2.16741683877804819444e-51, /* 0x3569F31D, 0x00000000 */
159
static int __kernel_rem_pio2(double[] x, ref double y_0_, ref double y_1_, ref double y_2_, int e0, int nx, int prec, int[] ipio2)
164
two24 = 1.67772160000000000000e+07, /* 0x41700000, 0x00000000 */
165
twon24 = 5.96046447753906250000e-08; /* 0x3E700000, 0x00000000 */
167
int jz,jx,jv,jp,jk,carry,n,i,j,k,m,q0,ih;
168
int[] iq = new int[20];
170
double[] f = new double[20];
171
double[] fq = new double[20];
172
double[] q = new double[20];
178
/* determine jx,jv,q0, note that 3>q0 */
180
jv = (e0-3)/24; if(jv<0) jv=0;
183
/* set up f[0] to f[jx+jk] where f[jx+jk] = ipio2[jv+jk] */
184
j = jv-jx; m = jx+jk;
185
for(i=0;i<=m;i++,j++) f[i] = (j<0)? zero : (double) ipio2[j];
187
/* compute q[0],q[1],...q[jk] */
188
for (i=0;i<=jk;i++) {
189
for(j=0,fw=0.0;j<=jx;j++) fw += x[j]*f[jx+i-j]; q[i] = fw;
194
/* distill q[] into iq[] reversingly */
195
for(i=0,j=jz,z=q[jz];j>0;i++,j--) {
196
fw = (double)((int)(twon24* z));
197
iq[i] = (int)(z-two24*fw);
202
z = scalbn(z,q0); /* actual value of z */
203
z -= 8.0*floor(z*0.125); /* trim off integer >= 8 */
207
if(q0>0) { /* need iq[jz-1] to determine n */
208
i = (iq[jz-1]>>(24-q0)); n += i;
209
iq[jz-1] -= i<<(24-q0);
210
ih = iq[jz-1]>>(23-q0);
212
else if(q0==0) ih = iq[jz-1]>>23;
213
else if(z>=0.5) ih=2;
215
if(ih>0) { /* q > 0.5 */
217
for(i=0;i<jz ;i++) { /* compute 1-q */
221
carry = 1; iq[i] = 0x1000000- j;
223
} else iq[i] = 0xffffff - j;
225
if(q0>0) { /* rare case: chance is 1 in 12 */
228
iq[jz-1] &= 0x7fffff; break;
230
iq[jz-1] &= 0x3fffff; break;
235
if(carry!=0) z -= scalbn(one,q0);
239
/* check if recomputation is needed */
242
for (i=jz-1;i>=jk;i--) j |= iq[i];
243
if(j==0) { /* need recomputation */
244
for(k=1;iq[jk-k]==0;k++); /* k = no. of terms needed */
246
for(i=jz+1;i<=jz+k;i++) { /* add q[jz+1] to q[jz+k] */
247
f[jx+i] = (double) ipio2[jv+i];
248
for(j=0,fw=0.0;j<=jx;j++) fw += x[j]*f[jx+i-j];
256
/* chop off zero terms */
259
while(iq[jz]==0) { jz--; q0-=24;}
260
} else { /* break z into 24-bit if necessary */
263
fw = (double)((int)(twon24*z));
264
iq[jz] = (int)(z-two24*fw);
267
} else iq[jz] = (int) z ;
270
/* convert integer "bit" chunk to floating-point value */
273
q[i] = fw*(double)iq[i]; fw*=twon24;
276
/* compute PIo2[0,...,jp]*q[jz,...,0] */
278
for(fw=0.0,k=0;k<=jp&&k<=jz-i;k++) fw += PIo2[k]*q[i+k];
282
/* compress fq[] into y[] */
286
for (i=jz;i>=0;i--) fw += fq[i];
287
y_0_ = (ih==0)? fw: -fw;
292
for (i=jz;i>=0;i--) fw += fq[i];
293
y_0_ = (ih==0)? fw: -fw;
295
for (i=1;i<=jz;i++) fw += fq[i];
296
y_1_ = (ih==0)? fw: -fw;
298
case 3: /* painful */
309
for (fw=0.0,i=jz;i>=2;i--) fw += fq[i];
311
y_0_ = fq[0]; y_1_ = fq[1]; y_2_ = fw;
313
y_0_ = -fq[0]; y_1_ = -fq[1]; y_2_ = -fw;