intervals 0.3.56
Sign up to get free protection for your applications and to get access to all the features.
- data/VERSION.txt +1 -0
- data/ext/crlibm/AUTHORS +2 -0
- data/ext/crlibm/COPYING +504 -0
- data/ext/crlibm/ChangeLog +80 -0
- data/ext/crlibm/INSTALL +182 -0
- data/ext/crlibm/Makefile.am +84 -0
- data/ext/crlibm/Makefile.in +530 -0
- data/ext/crlibm/NEWS +0 -0
- data/ext/crlibm/README +31 -0
- data/ext/crlibm/TODO +47 -0
- data/ext/crlibm/VERSION +1 -0
- data/ext/crlibm/aclocal.m4 +989 -0
- data/ext/crlibm/atan-itanium.c +846 -0
- data/ext/crlibm/atan-pentium.c +261 -0
- data/ext/crlibm/atan_accurate.c +244 -0
- data/ext/crlibm/atan_accurate.h +191 -0
- data/ext/crlibm/atan_fast.c +324 -0
- data/ext/crlibm/atan_fast.h +678 -0
- data/ext/crlibm/config.guess +1461 -0
- data/ext/crlibm/config.sub +1566 -0
- data/ext/crlibm/configure +7517 -0
- data/ext/crlibm/configure.ac +364 -0
- data/ext/crlibm/crlibm.h +125 -0
- data/ext/crlibm/crlibm_config.h +149 -0
- data/ext/crlibm/crlibm_config.h.in +148 -0
- data/ext/crlibm/crlibm_private.c +293 -0
- data/ext/crlibm/crlibm_private.h +658 -0
- data/ext/crlibm/csh_fast.c +631 -0
- data/ext/crlibm/csh_fast.h +771 -0
- data/ext/crlibm/double-extended.h +496 -0
- data/ext/crlibm/exp-td.c +962 -0
- data/ext/crlibm/exp-td.h +685 -0
- data/ext/crlibm/exp_accurate.c +197 -0
- data/ext/crlibm/exp_accurate.h +85 -0
- data/ext/crlibm/gappa/log-de-E0-logir0.gappa +106 -0
- data/ext/crlibm/gappa/log-de-E0.gappa +79 -0
- data/ext/crlibm/gappa/log-de.gappa +81 -0
- data/ext/crlibm/gappa/log-td-E0-logir0.gappa +126 -0
- data/ext/crlibm/gappa/log-td-E0.gappa +143 -0
- data/ext/crlibm/gappa/log-td-accurate-E0-logir0.gappa +230 -0
- data/ext/crlibm/gappa/log-td-accurate-E0.gappa +213 -0
- data/ext/crlibm/gappa/log-td-accurate.gappa +217 -0
- data/ext/crlibm/gappa/log-td.gappa +156 -0
- data/ext/crlibm/gappa/trigoSinCosCase3.gappa +204 -0
- data/ext/crlibm/gappa/trigoTanCase2.gappa +73 -0
- data/ext/crlibm/install-sh +269 -0
- data/ext/crlibm/log-de.c +431 -0
- data/ext/crlibm/log-de.h +732 -0
- data/ext/crlibm/log-td.c +852 -0
- data/ext/crlibm/log-td.h +819 -0
- data/ext/crlibm/log10-td.c +906 -0
- data/ext/crlibm/log10-td.h +823 -0
- data/ext/crlibm/log2-td.c +935 -0
- data/ext/crlibm/log2-td.h +821 -0
- data/ext/crlibm/maple/atan.mpl +359 -0
- data/ext/crlibm/maple/common-procedures.mpl +997 -0
- data/ext/crlibm/maple/csh.mpl +446 -0
- data/ext/crlibm/maple/double-extended.mpl +151 -0
- data/ext/crlibm/maple/exp-td.mpl +195 -0
- data/ext/crlibm/maple/log-de.mpl +243 -0
- data/ext/crlibm/maple/log-td.mpl +316 -0
- data/ext/crlibm/maple/log10-td.mpl +345 -0
- data/ext/crlibm/maple/log2-td.mpl +334 -0
- data/ext/crlibm/maple/trigo.mpl +728 -0
- data/ext/crlibm/maple/triple-double.mpl +58 -0
- data/ext/crlibm/missing +198 -0
- data/ext/crlibm/mkinstalldirs +40 -0
- data/ext/crlibm/rem_pio2_accurate.c +219 -0
- data/ext/crlibm/rem_pio2_accurate.h +53 -0
- data/ext/crlibm/scs_lib/AUTHORS +3 -0
- data/ext/crlibm/scs_lib/COPYING +504 -0
- data/ext/crlibm/scs_lib/ChangeLog +16 -0
- data/ext/crlibm/scs_lib/INSTALL +215 -0
- data/ext/crlibm/scs_lib/Makefile.am +18 -0
- data/ext/crlibm/scs_lib/Makefile.in +328 -0
- data/ext/crlibm/scs_lib/NEWS +0 -0
- data/ext/crlibm/scs_lib/README +9 -0
- data/ext/crlibm/scs_lib/TODO +4 -0
- data/ext/crlibm/scs_lib/addition_scs.c +623 -0
- data/ext/crlibm/scs_lib/config.guess +1461 -0
- data/ext/crlibm/scs_lib/config.sub +1566 -0
- data/ext/crlibm/scs_lib/configure +6226 -0
- data/ext/crlibm/scs_lib/division_scs.c +110 -0
- data/ext/crlibm/scs_lib/double2scs.c +174 -0
- data/ext/crlibm/scs_lib/install-sh +269 -0
- data/ext/crlibm/scs_lib/missing +198 -0
- data/ext/crlibm/scs_lib/mkinstalldirs +40 -0
- data/ext/crlibm/scs_lib/multiplication_scs.c +456 -0
- data/ext/crlibm/scs_lib/poly_fct.c +112 -0
- data/ext/crlibm/scs_lib/print_scs.c +73 -0
- data/ext/crlibm/scs_lib/rand_scs.c +63 -0
- data/ext/crlibm/scs_lib/scs.h +353 -0
- data/ext/crlibm/scs_lib/scs2double.c +391 -0
- data/ext/crlibm/scs_lib/scs2mpf.c +58 -0
- data/ext/crlibm/scs_lib/scs2mpfr.c +61 -0
- data/ext/crlibm/scs_lib/scs_private.c +23 -0
- data/ext/crlibm/scs_lib/scs_private.h +133 -0
- data/ext/crlibm/scs_lib/tests/tbx_timing.h +102 -0
- data/ext/crlibm/scs_lib/wrapper_scs.h +486 -0
- data/ext/crlibm/scs_lib/zero_scs.c +52 -0
- data/ext/crlibm/stamp-h.in +1 -0
- data/ext/crlibm/tests/Makefile.am +43 -0
- data/ext/crlibm/tests/Makefile.in +396 -0
- data/ext/crlibm/tests/blind_test.c +148 -0
- data/ext/crlibm/tests/generate_test_vectors.c +258 -0
- data/ext/crlibm/tests/soak_test.c +334 -0
- data/ext/crlibm/tests/test_common.c +627 -0
- data/ext/crlibm/tests/test_common.h +28 -0
- data/ext/crlibm/tests/test_perf.c +570 -0
- data/ext/crlibm/tests/test_val.c +249 -0
- data/ext/crlibm/trigo_accurate.c +500 -0
- data/ext/crlibm/trigo_accurate.h +331 -0
- data/ext/crlibm/trigo_fast.c +1219 -0
- data/ext/crlibm/trigo_fast.h +639 -0
- data/ext/crlibm/triple-double.h +878 -0
- data/ext/extconf.rb +31 -0
- data/ext/fpu.c +107 -0
- data/ext/jamis-mod.rb +591 -0
- data/lib/fpu.rb +287 -0
- data/lib/interval.rb +1170 -0
- data/lib/intervals.rb +212 -0
- data/lib/struct_float.rb +133 -0
- data/test/data_atan.txt +360 -0
- data/test/data_cos.txt +346 -0
- data/test/data_cosh.txt +3322 -0
- data/test/data_exp.txt +3322 -0
- data/test/data_log.txt +141 -0
- data/test/data_sin.txt +140 -0
- data/test/data_sinh.txt +3322 -0
- data/test/data_tan.txt +342 -0
- metadata +186 -0
@@ -0,0 +1,728 @@
|
|
1
|
+
|
2
|
+
Digits := 100:
|
3
|
+
interface(quiet=true):
|
4
|
+
read "common-procedures.mpl":
|
5
|
+
with(orthopoly):
|
6
|
+
|
7
|
+
|
8
|
+
mkdir("TEMPTRIG"):
|
9
|
+
|
10
|
+
|
11
|
+
|
12
|
+
# - Evaluation scheme :
|
13
|
+
# case 1 : return x
|
14
|
+
# case 2 (or Fast) : compute a simple polynomial
|
15
|
+
# case 3 : do an argument reduction...
|
16
|
+
|
17
|
+
|
18
|
+
|
19
|
+
########################################################
|
20
|
+
# Case 1 : Small arguments
|
21
|
+
# return x for sine and tan, return 1 for cos
|
22
|
+
########################################################
|
23
|
+
|
24
|
+
xmax_return_x_for_sin := 2^(-26):
|
25
|
+
xmax_return_1_for_cos_RN := sqrt(2^(-53)):
|
26
|
+
xmax_return_1_for_cos_RDIR:=2^(-26):
|
27
|
+
one_rounded_down := evalf(1-ulp(1/2)):
|
28
|
+
|
29
|
+
xmax_return_x_for_tan := 2^(-27):
|
30
|
+
|
31
|
+
|
32
|
+
|
33
|
+
|
34
|
+
########################################################
|
35
|
+
# Case 2 : simple polynomial approximation
|
36
|
+
########################################################
|
37
|
+
|
38
|
+
# We want to use the same polynomial in case 2 and 3.
|
39
|
+
# So see after arg red
|
40
|
+
|
41
|
+
#################################################
|
42
|
+
# Case 3 : Argument reduction
|
43
|
+
#################################################
|
44
|
+
|
45
|
+
|
46
|
+
|
47
|
+
#################################################
|
48
|
+
# CODY and WAITE Argument reduction
|
49
|
+
|
50
|
+
|
51
|
+
C := Pi/256:
|
52
|
+
invC:= nearest(1/C):
|
53
|
+
reminvC := evalf(1/C - invC):
|
54
|
+
expC:=ieeedouble(C)[2]:
|
55
|
+
epsinvC := abs(reminvC*C):
|
56
|
+
|
57
|
+
# There are three sets of constants :
|
58
|
+
# - split redC into two constants, for small values when we are concerned with absolute error
|
59
|
+
# - split redC into three constants, for larger values
|
60
|
+
# - split redC into three doubles, for the cases when we need
|
61
|
+
# good relative precision on the result and fear cancellation
|
62
|
+
|
63
|
+
|
64
|
+
|
65
|
+
|
66
|
+
# Fastest reduction using two-part Cody and Waite (up to |k|=2^22)
|
67
|
+
|
68
|
+
bitsCh_0:=34:
|
69
|
+
|
70
|
+
# 1/2 <= C/2^(expC+1) <1
|
71
|
+
Ch:= round(evalf( C * 2^(bitsCh_0-expC-1))) / (2^(bitsCh_0-expC-1)):
|
72
|
+
# recompute bitsCh in case we are lucky (and we are for bitsCh_0=32)
|
73
|
+
bitsCh:=1+log2(op(2,ieeedouble(Ch)[3])) : # this means the log of the denominator
|
74
|
+
|
75
|
+
Cl:=nearest(C - Ch):
|
76
|
+
# Cody and Waite argument reduction will work for |k|<kmax_cw2
|
77
|
+
kmax_cw2:=2^(53-bitsCh):
|
78
|
+
|
79
|
+
# The constants to move to the .h file
|
80
|
+
RR_CW2_CH := Ch:
|
81
|
+
RR_CW2_MCL := -Cl:
|
82
|
+
XMAX_CODY_WAITE_2 := nearest(kmax_cw2*C):
|
83
|
+
|
84
|
+
# The error in this case (we need absolute error)
|
85
|
+
delta_repr_C_cw2 := abs(C-Ch-Cl):
|
86
|
+
delta_round_cw2 := kmax_cw2* 1/2 * ulp(Cl) :
|
87
|
+
delta_cody_waite_2 := kmax_cw2 * delta_repr_C_cw2 + delta_round_cw2:
|
88
|
+
# This is the delta on y, the reduced argument
|
89
|
+
|
90
|
+
#log2(%);
|
91
|
+
|
92
|
+
|
93
|
+
|
94
|
+
# Slower reduction using three-part Cody and Waite, up to |k|=2^31
|
95
|
+
|
96
|
+
bitsCh_0:=23: # 22 or 23
|
97
|
+
Ch:= round(evalf( C * 2^(bitsCh_0-expC-1))) / (2^(bitsCh_0-expC-1)):
|
98
|
+
# recompute bitsCh in case we are lucky
|
99
|
+
bitsCh:=1+log2(op(2,ieeedouble(Ch)[3])) : # this means the log of the denominator
|
100
|
+
|
101
|
+
r := C-Ch:
|
102
|
+
Cmed := round(evalf( r * 2^(2*bitsCh-expC-1))) / (2^(2*bitsCh-expC-1)):
|
103
|
+
bitsCmed:=1+log2(op(2,ieeedouble(Cmed)[3])) :
|
104
|
+
|
105
|
+
Cl:=nearest(C - Ch - Cmed):
|
106
|
+
|
107
|
+
kmax_cw3 := 2^min(53-bitsCh, 53-bitsCmed, 31):# Otherwise we have integer overflow
|
108
|
+
|
109
|
+
|
110
|
+
|
111
|
+
# The constants to move to the .h file
|
112
|
+
RR_CW3_CH := Ch:
|
113
|
+
RR_CW3_CM := Cmed:
|
114
|
+
RR_CW3_MCL := -Cl:
|
115
|
+
XMAX_CODY_WAITE_3 := nearest(kmax_cw3*C):
|
116
|
+
|
117
|
+
# The error in this case (we need absolute error)
|
118
|
+
delta_repr_C_cw3 := abs(C - Ch - Cmed - Cl):
|
119
|
+
delta_round_cw3 := kmax_cw3 * 1/2 * ulp(Cl) :
|
120
|
+
delta_cody_waite_3 := kmax_cw3 * delta_repr_C_cw3 + delta_round_cw3:
|
121
|
+
# This is the delta on y, the reduced argument
|
122
|
+
|
123
|
+
#log2(%);
|
124
|
+
|
125
|
+
|
126
|
+
|
127
|
+
|
128
|
+
|
129
|
+
# Third range reduction, using double-double arithmetic, for |k| up to 2^51-1
|
130
|
+
|
131
|
+
# This max int value can be produced by DOUBLE2LONGINT
|
132
|
+
kmax:=2^46-1:
|
133
|
+
XMAX_DDRR:=nearest(kmax*C):
|
134
|
+
|
135
|
+
#in this case we have C stored as 3 doubles
|
136
|
+
Ch := nearest(C):
|
137
|
+
Cmed := nearest(C-Ch):
|
138
|
+
Cl := nearest(C-Ch-Cmed):
|
139
|
+
|
140
|
+
RR_DD_MCH := -Ch:
|
141
|
+
RR_DD_MCM := -Cmed:
|
142
|
+
RR_DD_CL := Cl:
|
143
|
+
|
144
|
+
delta_repr_C := abs(C - Ch - Cmed - Cl):
|
145
|
+
|
146
|
+
# and we have only exact Add12 and Mul12 operations. The only place
|
147
|
+
# with possible rounding errors is:
|
148
|
+
# Add22 (pyh, pyl, (x + kch_h) , (kcm_l - kd*RR_DD_CL), th, tl) ;
|
149
|
+
# where (x + kch_h) is exact (Sterbenz) with up to kmax bits of cancellation
|
150
|
+
# and the error is simply the error in (kcm_l - kd*RR_DD_CL)
|
151
|
+
# At the very worst :
|
152
|
+
delta_round :=
|
153
|
+
kmax * 1/2 * ulp(Cl) # for kd*RR_DD_CL
|
154
|
+
+ kmax*ulp(Cl) # for the subtraction
|
155
|
+
+ 2^(-100) * Pi/512 : # for the Add22
|
156
|
+
delta_RR_DD := kmax * delta_repr_C + delta_round:
|
157
|
+
|
158
|
+
# the last case, Payne and Hanek reduction, gives a very small delta:
|
159
|
+
# red arg is on 9*30 bits, then rounded to a double-double (106 bits)
|
160
|
+
# This should, of course, be optimized some day
|
161
|
+
delta_PayneHanek := 2^(-100):
|
162
|
+
|
163
|
+
# Finally the max delta on the reduced argument is
|
164
|
+
delta_ArgRed := max(delta_cody_waite_2, delta_cody_waite_3,
|
165
|
+
delta_RR_DD, delta_PayneHanek):
|
166
|
+
|
167
|
+
#print("delta_ArgRed to move to the .gappa file = ", evalf(delta_ArgRed)):
|
168
|
+
#log2(delta_ArgRed);
|
169
|
+
|
170
|
+
|
171
|
+
|
172
|
+
|
173
|
+
|
174
|
+
# Now we use the above absolute error when k mod 256 <> 0
|
175
|
+
# otherwise we need to worry about relative accuracy of the result.
|
176
|
+
|
177
|
+
# First, what is the worst case for cancellation ?
|
178
|
+
|
179
|
+
emax := ieeedouble(XMAX_DDRR)[2] +1 :
|
180
|
+
# above emax, we will use Payne and Hanek so we do not worry
|
181
|
+
|
182
|
+
(wcn, wce, wceps) := WorstCaseForAdditiveRangeReduction(2,53,-8, emax, C):
|
183
|
+
wcx := wcn * 2^wce:
|
184
|
+
wck := round(wcx/C):
|
185
|
+
wcy := wcx - wck*C:
|
186
|
+
|
187
|
+
#log2(wcy); # y > 2^(-67);
|
188
|
+
|
189
|
+
# In these cases we use the double-double range reduction, for |k|<kmax_cw3
|
190
|
+
# and the relative precision in the worst case is for wcy
|
191
|
+
|
192
|
+
delta_round := kmax_cw3 * 1/2 * ulp(Cl) # for kd*RR_DD_CL
|
193
|
+
+ kmax_cw3 * ulp(Cl) : # for the subtraction
|
194
|
+
|
195
|
+
delta_RR_DD := kmax_cw3 * delta_repr_C + delta_round:
|
196
|
+
|
197
|
+
eps_ArgRed := (1+delta_RR_DD/wcy)*(1+2^(-100)) -1:
|
198
|
+
|
199
|
+
#log2(eps_ArgRed);
|
200
|
+
|
201
|
+
# In all the other cases we use Payne and Hanek, and eps_ArgRed is
|
202
|
+
# much smaller, so this is the max.
|
203
|
+
|
204
|
+
|
205
|
+
|
206
|
+
###########
|
207
|
+
# Polynomials for do_sine and do_cos, and for the case 2
|
208
|
+
degreeSin := 8:
|
209
|
+
degreeCos := 7:
|
210
|
+
|
211
|
+
maxepsk := (1+epsinvC)*(1+2^(-53))-1:
|
212
|
+
|
213
|
+
ymaxCase3 := evalf(Pi/512 + XMAX_DDRR*maxepsk):
|
214
|
+
#print("ymaxCase3 to move to the .gappa file = ", ymaxCase3):
|
215
|
+
|
216
|
+
|
217
|
+
y2maxCase3 := ymaxCase3^2:
|
218
|
+
# These are the parameters that can be varied to fine-tune performance
|
219
|
+
# (they should always be larger than Pi/512
|
220
|
+
xmaxCosCase2 := Pi/256:
|
221
|
+
xmaxSinCase2 := Pi/256:
|
222
|
+
|
223
|
+
|
224
|
+
x2maxSinCase2:= xmaxSinCase2^2:
|
225
|
+
x2maxCosCase2:= xmaxCosCase2^2:
|
226
|
+
|
227
|
+
# We had the difficulty here to find minimax polynomials which are good for case 2
|
228
|
+
# as well as for case 3. A simple solution was to set xmaxCosCase2 =ymaxCase3...
|
229
|
+
# However we found another answer: in the future we intend to use these polynomials for second
|
230
|
+
# step, too. Therefore, no minimax, only Taylor.
|
231
|
+
|
232
|
+
polySin:= poly_exact(convert( series(sin(x), x=0, degreeSin+1), polynom)):
|
233
|
+
polyTs := expand(polySin/x-1):
|
234
|
+
polyTs2 := subs(x=sqrt(y), polyTs):
|
235
|
+
|
236
|
+
polyCos := poly_exact (convert( series(cos(x), x=0, degreeCos+1), polynom)):
|
237
|
+
polyTc2 := subs(x=sqrt(y), polyCos - 1):
|
238
|
+
|
239
|
+
epsApproxSinCase2 := numapprox[infnorm]((x*polyTs+x -sin(x))/sin(x), x=0..xmaxSinCase2):
|
240
|
+
epsApproxSinCase3 := numapprox[infnorm]((x*polyTs +x -sin(x))/sin(x), x=0..ymaxCase3):
|
241
|
+
|
242
|
+
deltaApproxSinCase2 := numapprox[infnorm]((x*polyTs+x -sin(x)), x=0..xmaxSinCase2):
|
243
|
+
deltaApproxSinCase3 := numapprox[infnorm]((x*polyTs +x -sin(x)), x=0..ymaxCase3):
|
244
|
+
|
245
|
+
deltaApproxCosCase2:= numapprox[infnorm](polyCos - cos(x), x=0..xmaxCosCase2):
|
246
|
+
deltaApproxCosCase3:= numapprox[infnorm](polyCos - cos(x), x=0..ymaxCase3):
|
247
|
+
|
248
|
+
#print("deltaApproxSinCase3 to move to the .gappa file = ", deltaApproxSinCase3):
|
249
|
+
#print("deltaApproxCosCase3 to move to the .gappa file = ", deltaApproxCosCase3):
|
250
|
+
|
251
|
+
|
252
|
+
########################## Case 2 for sine ###########################
|
253
|
+
|
254
|
+
|
255
|
+
|
256
|
+
# evaluate this polynomial in double. The error on x*x is at most half an ulp
|
257
|
+
errlist:=errlist_quickphase_horner(degree(polyTs2),0,0,2^(-53), 0):
|
258
|
+
(eps_rounding_Ts, delta_rounding_Ts, minTs, maxTs):=
|
259
|
+
compute_horner_rounding_error(polyTs2,y,x2maxSinCase2, errlist, true):
|
260
|
+
|
261
|
+
eps_poly_TsCase2 := numapprox[infnorm]((x*polyTs)/(sin(x)-x) -1, x=0..xmaxSinCase2):
|
262
|
+
maxeps2 := (1+eps_poly_TsCase2)*(1+eps_rounding_Ts)*(1+2^(-53))-1:
|
263
|
+
|
264
|
+
maxepstotalSinCase2 := maxeps2 * numapprox[infnorm](1-x/sin(x), x=0..xmaxSinCase2):
|
265
|
+
rnconstantSinCase2 := evalf(compute_rn_constant(maxepstotalSinCase2)):
|
266
|
+
|
267
|
+
|
268
|
+
|
269
|
+
|
270
|
+
|
271
|
+
##################################### Case2 cos ###########################
|
272
|
+
|
273
|
+
# evaluate this polynomial in double. The error on x*x is at most half an ulp
|
274
|
+
errlist := errlist_quickphase_horner(degree(polyTc2),0,0,2**(-53), 0):
|
275
|
+
(eps_rounding_Tc, delta_rounding_Tc, minTc, maxTc):=
|
276
|
+
compute_horner_rounding_error(polyTc2,y,x2maxCosCase2, errlist, true):
|
277
|
+
|
278
|
+
# Then we have an Add12 which is exact. The result is greater then cos(xmaxCosCase2):
|
279
|
+
miny := cos(xmaxCosCase2):
|
280
|
+
maxepstotalCosCase2 := (delta_rounding_Tc + deltaApproxCosCase2) / miny :
|
281
|
+
#log2(%);
|
282
|
+
rnconstantCosCase2 := evalf(compute_rn_constant(maxepstotalCosCase2)):
|
283
|
+
|
284
|
+
|
285
|
+
|
286
|
+
|
287
|
+
######################## Case2 Tangent #########################
|
288
|
+
#
|
289
|
+
# Compute the Taylor series
|
290
|
+
degreeTanCase2 := 12:
|
291
|
+
xmaxTanCase2 := 2**(-4):
|
292
|
+
xminTanCase2 := 2**(-30):
|
293
|
+
|
294
|
+
Poly_P := convert(series(tan(sqrt(x))/(x^(3/2))-1/x, x=0, degreeTanCase2*4),polynom):
|
295
|
+
Poly_cheb := numapprox[chebpade](Poly_P, x=xminTanCase2..xmaxTanCase2^2, [degreeTanCase2/2-2,0]):
|
296
|
+
polyTanCase2 := poly_exact2(expand(x + x^3 * subs(x=x^2, Poly_cheb)), 4):
|
297
|
+
|
298
|
+
#polyTanCase2 := poly_exact2(convert(series(tan(x), x=0, degreeTanCase2), polynom), 4):
|
299
|
+
|
300
|
+
maxepsApproxTanCase2:=numapprox[infnorm](1 - polyTanCase2 / tan(x), x=xminTanCase2..xmaxTanCase2):
|
301
|
+
|
302
|
+
maxepsOverXTanCase2 :=numapprox[infnorm]((1 - polyTanCase2 / tan(x))/x, x=xminTanCase2..xmaxTanCase2):
|
303
|
+
# Now we pass these values to Gappa
|
304
|
+
|
305
|
+
filename:="TEMPTRIG/TanCase2.sed":
|
306
|
+
fd:=fopen(filename, WRITE, TEXT):
|
307
|
+
t3h, t3l := hi_lo(coeff(polyTanCase2,x,3)):
|
308
|
+
fprintf(fd, " s/_t3h/%1.40e/g\n", t3h):
|
309
|
+
fprintf(fd, " s/_t3l/%1.40e/g\n", t3l):
|
310
|
+
fprintf(fd, " s/_t5/%1.40e/g\n", coeff(polyTanCase2,x,5)):
|
311
|
+
fprintf(fd, " s/_t7/%1.40e/g\n", coeff(polyTanCase2,x,7)):
|
312
|
+
fprintf(fd, " s/_t9/%1.40e/g\n", coeff(polyTanCase2,x,9)):
|
313
|
+
fprintf(fd, " s/_t11/%1.40e/g\n", coeff(polyTanCase2,x,11)):
|
314
|
+
fprintf(fd, " s/_xmax/%1.40e/g\n", xmaxTanCase2):
|
315
|
+
fprintf(fd, " s/_maxEpsApproxOverX/%1.40e/g\n", maxepsOverXTanCase2*1.00001):
|
316
|
+
fprintf(fd, " s/_maxEpsApprox/%1.40e/g\n", maxepsApproxTanCase2*1.00001):
|
317
|
+
fclose(fd):
|
318
|
+
|
319
|
+
printf("\n\n************ DONE TEMPTRIG/TanCase2.sed ************\n");
|
320
|
+
printf("Now you should use \n sed -f TEMPTRIG/TanCase2.sed trigoTanCase2.gappa | gappa > /dev/null \n");
|
321
|
+
|
322
|
+
|
323
|
+
|
324
|
+
maxepstotalTanCase2:=4.59602e-19: # Cut from Gappa output
|
325
|
+
|
326
|
+
log2(maxepstotalTanCase2): # almost 61 bits
|
327
|
+
|
328
|
+
|
329
|
+
|
330
|
+
|
331
|
+
|
332
|
+
|
333
|
+
###############################################################################
|
334
|
+
# Computing errors for Case3 : now we have an error due to arg red
|
335
|
+
|
336
|
+
# First DoSinZero. The notations are those of the paper proof
|
337
|
+
|
338
|
+
# Approximation error already computed above as epsApproxSinCase3;
|
339
|
+
|
340
|
+
# polynomial evaluation in double, with an error on y*y of epsilonArgRed
|
341
|
+
errlist:=errlist_quickphase_horner(degree(polyTs2),0,0,eps_ArgRed, 0):
|
342
|
+
(epsRoundingTsSinZero, deltaRoundingTsSinZero, minTs, maxTs):=
|
343
|
+
compute_horner_rounding_error(polyTs2,y,y2maxCase3, errlist, true):
|
344
|
+
|
345
|
+
# just as in the paper proof
|
346
|
+
maxepsSinZero1 := (1+epsApproxSinCase3)*(1+epsRoundingTsSinZero)*(1+2^(-53))*(1+2^(-53)) - 1:
|
347
|
+
|
348
|
+
# just as in the paper proof. For x>0 the absolute values are as given
|
349
|
+
|
350
|
+
epstotalSinZero := ( (x-sin(x))*maxepsSinZero1 + x*eps_ArgRed + 2^(-53)*x^3/3 ) / sin(x):
|
351
|
+
maxepstotalSinZero := numapprox[infnorm]( epstotalSinZero , x=2^(-30)..ymaxCase3):
|
352
|
+
|
353
|
+
printf("\nMax rel error for DoSinZero is %1.5e, if it's smaller than 2^(-66) (%1.5e) then the proof is OK\n\n", maxepstotalSinZero, 2^(-66)):
|
354
|
+
|
355
|
+
|
356
|
+
|
357
|
+
|
358
|
+
|
359
|
+
##############################SinCosCase3############################
|
360
|
+
SinCosSize:= 128: # size f the table
|
361
|
+
|
362
|
+
|
363
|
+
# The Gappa files in TEMPTRIG
|
364
|
+
for i from 1 to SinCosSize/2 do
|
365
|
+
filename:=cat("TEMPTRIG/SinACosA_",i,".sed"):
|
366
|
+
fd:=fopen(filename, WRITE, TEXT):
|
367
|
+
|
368
|
+
# The table values
|
369
|
+
s:=hi_lo(sin(i*Pi/(2*SinCosSize))):
|
370
|
+
c:=hi_lo(cos(i*Pi/(2*SinCosSize))):
|
371
|
+
fprintf(fd, " s/_cah/%1.40e/g\n", c[1]):
|
372
|
+
fprintf(fd, " s/_cal/%1.40e/g\n", c[2]):
|
373
|
+
fprintf(fd, " s/_sah/%1.40e/g\n", s[1]):
|
374
|
+
fprintf(fd, " s/_sal/%1.40e/g\n", s[2]):
|
375
|
+
|
376
|
+
# The polynomial coefficients
|
377
|
+
fprintf(fd, " s/_s3/%1.40e/g\n", coeff(polySin,x,3)):
|
378
|
+
fprintf(fd, " s/_s5/%1.40e/g\n", coeff(polySin,x,5)):
|
379
|
+
fprintf(fd, " s/_s7/%1.40e/g\n", coeff(polySin,x,7)):
|
380
|
+
fprintf(fd, " s/_s9/%1.40e/g\n", coeff(polySin,x,9)):
|
381
|
+
fprintf(fd, " s/_c2/%1.40e/g\n", coeff(polyCos,x,2)):
|
382
|
+
fprintf(fd, " s/_c4/%1.40e/g\n", coeff(polyCos,x,4)):
|
383
|
+
fprintf(fd, " s/_c6/%1.40e/g\n", coeff(polyCos,x,6)):
|
384
|
+
fprintf(fd, " s/_c8/%1.40e/g\n", coeff(polyCos,x,8)):
|
385
|
+
|
386
|
+
# The approximation errors
|
387
|
+
fprintf(fd, " s/_ymaxCase3/%1.40e/g\n", ymaxCase3*1.00001):
|
388
|
+
fprintf(fd, " s/_delta_ArgRed/%1.40e/g\n", delta_ArgRed*1.00001):
|
389
|
+
fprintf(fd, " s/_delta_approx_Sin_Case3/%1.40e/g\n", deltaApproxSinCase3*1.00001):
|
390
|
+
fprintf(fd, " s/_delta_approx_Cos_Case3/%1.40e/g\n", deltaApproxCosCase3*1.00001):
|
391
|
+
|
392
|
+
fclose(fd):
|
393
|
+
od:
|
394
|
+
|
395
|
+
|
396
|
+
printf("************ DONE TEMPTRIG/*.sed ************\n"):
|
397
|
+
|
398
|
+
# A shell script to use them
|
399
|
+
filename:="trigo_test.sh":
|
400
|
+
fd:=fopen(filename, WRITE, TEXT):
|
401
|
+
fprintf(fd, "#!/bin/sh\n"):
|
402
|
+
fprintf(fd, "# You probably need to edit the path to the gappa executable\n"):
|
403
|
+
fprintf(fd, "for file in TEMPTRIG/SinACosA*.sed \n"):
|
404
|
+
fprintf(fd, "do\n"):
|
405
|
+
fprintf(fd, " echo $file:\n"):
|
406
|
+
fprintf(fd, " sed -f $file trigoSinCosCase3.gappa | ~/gappa/src/gappa > /dev/null\n"):
|
407
|
+
fprintf(fd, " echo\n"):
|
408
|
+
fprintf(fd, "done\n"):
|
409
|
+
fclose(fd):
|
410
|
+
|
411
|
+
printf("************ DONE trigo_test.sh ************\n"):
|
412
|
+
printf("Now you should run\n"):
|
413
|
+
printf(" sh trigo_test.sh 2>TEMPTRIG/Gappa.out\n"):
|
414
|
+
|
415
|
+
printf("Then look at TEMPTRIG/Gappa.out. It shouldn't contain 'No proof'.\n This means that everything is OK and the rounding constants in TEMPTRIG/trigo_fast.h are proven upper bounds.\n\n"):
|
416
|
+
|
417
|
+
|
418
|
+
|
419
|
+
|
420
|
+
# This value has been validated by Gappa (using all the previous)
|
421
|
+
maxepstotalSinCosCase3:=3*2^(-66):
|
422
|
+
rnconstantSinCosCase3 := evalf(compute_rn_constant(maxepstotalSinCosCase3)):
|
423
|
+
|
424
|
+
|
425
|
+
# The error of sin, the error of cos, then the error of Div22
|
426
|
+
maxepstotalTanCase3:= 2.1*maxepstotalSinCosCase3:
|
427
|
+
rnconstantTanCase3 := evalf(compute_rn_constant(maxepstotalTanCase3)):
|
428
|
+
|
429
|
+
|
430
|
+
|
431
|
+
|
432
|
+
|
433
|
+
##############################################
|
434
|
+
## Compute constants for SCS arg red
|
435
|
+
oldDigits:=Digits:
|
436
|
+
Digits:=1000:
|
437
|
+
# for 2/Pi:
|
438
|
+
n:=round(2^(30*48)*evalf(2/Pi)):
|
439
|
+
digitlist:=[]:
|
440
|
+
for i from 1 to 48 do
|
441
|
+
r:=n mod (2^30):
|
442
|
+
n:=floor(n/(2^30)):
|
443
|
+
hexstring:= convert(convert(r,hex),string):
|
444
|
+
digitlist:=[hexstring, op(digitlist)]:
|
445
|
+
end:
|
446
|
+
digitlist:
|
447
|
+
|
448
|
+
# for 256/Pi:
|
449
|
+
n:=round(2^(30*47)*evalf(256/Pi)):
|
450
|
+
digitlist:=[]:
|
451
|
+
for i from 1 to 48 do
|
452
|
+
r:=n mod (2^30):
|
453
|
+
n:=floor(n/(2^30)):
|
454
|
+
hexstring:= convert(convert(r,hex),string):
|
455
|
+
digitlist:=[hexstring, op(digitlist)]:
|
456
|
+
end:
|
457
|
+
digitlist:
|
458
|
+
Digits:=oldDigits:
|
459
|
+
|
460
|
+
|
461
|
+
|
462
|
+
# an auxiliary output function:
|
463
|
+
# Outputs the high part of a double, and the double in comment.
|
464
|
+
# As all these high parts are used in code as
|
465
|
+
# if(absxhi < XMAX_COS_CASE2)
|
466
|
+
# we have to remove one LSB to the high part, or, divide var by
|
467
|
+
# (1+2^(-20))
|
468
|
+
# Now we have absxhi<highpart(var/(1+2^(-20))
|
469
|
+
# => absxhi*(1+2^(-20))<var
|
470
|
+
# => x<var
|
471
|
+
outputHighPart:=proc(cvarname, var)
|
472
|
+
local varMinusLSB:
|
473
|
+
Digits:=8:
|
474
|
+
varMinusLSB:=var/(1+2^(-20)):
|
475
|
+
("#define " || cvarname || " 0x" || (ieeehexa(varMinusLSB)[1])
|
476
|
+
|| " /* " || (convert(evalf(varMinusLSB),string)) || " */" )
|
477
|
+
end proc:
|
478
|
+
|
479
|
+
|
480
|
+
|
481
|
+
|
482
|
+
|
483
|
+
# Output:
|
484
|
+
|
485
|
+
filename:="TEMPTRIG/trigo_fast.h":
|
486
|
+
fd:=fopen(filename, WRITE, TEXT):
|
487
|
+
|
488
|
+
fprintf(fd, "#include \"crlibm.h\"\n#include \"crlibm_private.h\"\n"):
|
489
|
+
|
490
|
+
fprintf(fd, "\n/*File generated by maple/trigo.pl*/\n"):
|
491
|
+
fprintf(fd, "\n"):
|
492
|
+
|
493
|
+
|
494
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_RETURN_X_FOR_SIN", xmax_return_x_for_sin) ):
|
495
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_SIN_CASE2 ", xmaxSinCase2) ):
|
496
|
+
|
497
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_RETURN_1_FOR_COS_RN", xmax_return_1_for_cos_RN) ):
|
498
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_RETURN_1_FOR_COS_RDIR", xmax_return_1_for_cos_RDIR) ):
|
499
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_COS_CASE2 ", xmaxCosCase2) ):
|
500
|
+
|
501
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_RETURN_X_FOR_TAN", xmax_return_x_for_tan) ):
|
502
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_TAN_CASE2 ", xmaxTanCase2) ):
|
503
|
+
|
504
|
+
fprintf(fd, "\n"):
|
505
|
+
fprintf(fd, "#define ONE_ROUNDED_DOWN %1.25e \n", one_rounded_down):
|
506
|
+
fprintf(fd, "\n"):
|
507
|
+
|
508
|
+
fprintf(fd, "#define EPS_SIN_CASE2 %1.25e \n", maxepstotalSinCase2):
|
509
|
+
fprintf(fd, "#define RN_CST_SIN_CASE2 %1.25f \n", rnconstantSinCase2):
|
510
|
+
fprintf(fd, "\n"):
|
511
|
+
fprintf(fd, "#define EPS_COS_CASE2 %1.25e \n", maxepstotalCosCase2):
|
512
|
+
fprintf(fd, "#define RN_CST_COS_CASE2 %1.25f \n", rnconstantCosCase2):
|
513
|
+
fprintf(fd, "\n"):
|
514
|
+
fprintf(fd, "#define EPS_SINCOS_CASE3 %1.25e \n", maxepstotalSinCosCase3):
|
515
|
+
fprintf(fd, "#define RN_CST_SINCOS_CASE3 %1.25f \n", rnconstantSinCosCase3):
|
516
|
+
fprintf(fd, "\n"):
|
517
|
+
fprintf(fd, "#define EPS_TAN_CASE2 %1.25e \n", maxepstotalTanCase2):
|
518
|
+
fprintf(fd, "#define EPS_TAN_CASE3 %1.25e \n", maxepstotalTanCase3):
|
519
|
+
fprintf(fd, "#define RN_CST_TAN_CASE3 %1.25f \n", rnconstantTanCase3):
|
520
|
+
|
521
|
+
fprintf(fd, "\n"):
|
522
|
+
|
523
|
+
fprintf(fd, "#define INV_PIO256 %1.25f \n", 1/C):
|
524
|
+
fprintf(fd, "\n"):
|
525
|
+
|
526
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_CODY_WAITE_2", XMAX_CODY_WAITE_2) ):
|
527
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_CODY_WAITE_3", XMAX_CODY_WAITE_3) ):
|
528
|
+
fprintf(fd, "%s\n", outputHighPart("XMAX_DDRR ", XMAX_DDRR) ):
|
529
|
+
#fprintf(fd, "%s\n", outputHighPart("", ) ):
|
530
|
+
fprintf(fd, "\n"):
|
531
|
+
|
532
|
+
fprintf(fd, "#define RR_CW2_CH %1.25e\n", RR_CW2_CH):
|
533
|
+
fprintf(fd, "#define RR_CW2_MCL %1.25e\n", RR_CW2_MCL):
|
534
|
+
fprintf(fd, "\n"):
|
535
|
+
|
536
|
+
fprintf(fd, "#define RR_CW3_CH %1.25e\n", RR_CW3_CH):
|
537
|
+
fprintf(fd, "#define RR_CW3_CM %1.25e\n", RR_CW3_CM):
|
538
|
+
fprintf(fd, "#define RR_CW3_MCL %1.25e\n", RR_CW3_MCL):
|
539
|
+
fprintf(fd, "\n"):
|
540
|
+
|
541
|
+
fprintf(fd, "#define RR_DD_MCH %1.25e\n", RR_DD_MCH):
|
542
|
+
fprintf(fd, "#define RR_DD_MCM %1.25e\n", RR_DD_MCM):
|
543
|
+
fprintf(fd, "#define RR_DD_CL %1.25e\n", RR_DD_CL):
|
544
|
+
fprintf(fd, "\n"):
|
545
|
+
|
546
|
+
fprintf(fd, "\n"):
|
547
|
+
|
548
|
+
fprintf(fd, "\n\n"):
|
549
|
+
|
550
|
+
# The 256/Pi SCS array
|
551
|
+
fprintf(fd, "static const int digits_256_over_pi[] = \n{"):
|
552
|
+
for i from 0 to 11 do
|
553
|
+
for j from 1 to 4 do
|
554
|
+
fprintf(fd, " 0x%s, \t",digitlist[4*i+j]):
|
555
|
+
end:
|
556
|
+
fprintf(fd, "\n "):
|
557
|
+
end:
|
558
|
+
fprintf(fd, "};\n\n"):
|
559
|
+
|
560
|
+
# The Pi/256 SCS constant
|
561
|
+
fprintf(fd, "static const scs Pio256=\n"):
|
562
|
+
WriteSCS(fd, evalf(C)):
|
563
|
+
fprintf(fd, ";\n#define Pio256_ptr (scs_ptr)(& Pio256)\n\n"):
|
564
|
+
|
565
|
+
fprintf(fd,"#ifdef WORDS_BIGENDIAN\n"):
|
566
|
+
for isbig from 1 to 0 by -1 do
|
567
|
+
|
568
|
+
if(isbig=0) then
|
569
|
+
fprintf(fd,"#else\n"):
|
570
|
+
fi:
|
571
|
+
|
572
|
+
# The sine polynomial
|
573
|
+
|
574
|
+
fprintf(fd, "static db_number const s3 = "):
|
575
|
+
printendian(fd, coeff(polySin,x,3), isbig):
|
576
|
+
fprintf(fd, ";\n"):
|
577
|
+
fprintf(fd, "static db_number const s5 = "):
|
578
|
+
printendian(fd, coeff(polySin,x,5), isbig):
|
579
|
+
fprintf(fd, ";\n"):
|
580
|
+
fprintf(fd, "static db_number const s7 = "):
|
581
|
+
printendian(fd, coeff(polySin,x,7), isbig):
|
582
|
+
fprintf(fd, ";\n\n"):
|
583
|
+
|
584
|
+
|
585
|
+
# the cos polynomial
|
586
|
+
|
587
|
+
fprintf(fd, "static db_number const c2 = "):
|
588
|
+
printendian(fd, coeff(polyCos,x,2), isbig):
|
589
|
+
fprintf(fd, ";\n"):
|
590
|
+
fprintf(fd, "static db_number const c4 = "):
|
591
|
+
printendian(fd, coeff(polyCos,x,4), isbig):
|
592
|
+
fprintf(fd, ";\n"):
|
593
|
+
fprintf(fd, "static db_number const c6 = "):
|
594
|
+
printendian(fd, coeff(polyCos,x,6), isbig):
|
595
|
+
fprintf(fd, ";\n\n"):
|
596
|
+
|
597
|
+
|
598
|
+
# the tan polynomial
|
599
|
+
|
600
|
+
t3h, t3l := hi_lo(coeff(polyTanCase2,x,3)):
|
601
|
+
fprintf(fd, "static db_number const t3h = "):
|
602
|
+
printendian(fd, t3h, isbig):
|
603
|
+
fprintf(fd, ";\n"):
|
604
|
+
fprintf(fd, "static db_number const t3l = "):
|
605
|
+
printendian(fd, t3l, isbig):
|
606
|
+
fprintf(fd, ";\n"):
|
607
|
+
fprintf(fd, "static db_number const t5 = "):
|
608
|
+
printendian(fd, coeff(polyTanCase2,x,5), isbig):
|
609
|
+
fprintf(fd, ";\n"):
|
610
|
+
fprintf(fd, "static db_number const t7 = "):
|
611
|
+
printendian(fd, coeff(polyTanCase2,x,7), isbig):
|
612
|
+
fprintf(fd, ";\n"):
|
613
|
+
fprintf(fd, "static db_number const t9 = "):
|
614
|
+
printendian(fd, coeff(polyTanCase2,x,9), isbig):
|
615
|
+
fprintf(fd, ";\n"):
|
616
|
+
fprintf(fd, "static db_number const t11 = "):
|
617
|
+
printendian(fd, coeff(polyTanCase2,x,11), isbig):
|
618
|
+
fprintf(fd, ";\n\n"):
|
619
|
+
|
620
|
+
|
621
|
+
# The sincos table
|
622
|
+
|
623
|
+
fprintf(fd, "\n/* sine and cos of kPi/256 in double-double */\n"):
|
624
|
+
fprintf(fd, "static db_number const sincosTable[%d] =\n{\n", 4*(SinCosSize/2+1)):
|
625
|
+
for i from 0 to SinCosSize/2 do
|
626
|
+
s:=hi_lo(sin(i*Pi/(2*SinCosSize))):
|
627
|
+
c:=hi_lo(cos(i*Pi/(2*SinCosSize))):
|
628
|
+
printendian(fd,s[1],isbig):
|
629
|
+
fprintf(fd," ,\n"):
|
630
|
+
printendian(fd,s[2],isbig):
|
631
|
+
fprintf(fd," ,\n"):
|
632
|
+
printendian(fd,c[1],isbig):
|
633
|
+
fprintf(fd," ,\n"):
|
634
|
+
printendian(fd,c[2],isbig):
|
635
|
+
if i<SinCosSize-1 then fprintf(fd," ,\n"): fi:
|
636
|
+
od:
|
637
|
+
fprintf(fd, "\n};\n\n"):
|
638
|
+
|
639
|
+
|
640
|
+
od:
|
641
|
+
fprintf(fd,"#endif /* WORDS_BIGENDIAN */\n\n\n"):
|
642
|
+
|
643
|
+
fclose(fd):
|
644
|
+
|
645
|
+
printf("\n\n************ DONE TEMPTRIG/trigo_fast.h ************\n Copy it to the crlibm source directory.\n\n");
|
646
|
+
|
647
|
+
|
648
|
+
#################################################
|
649
|
+
# Stuff for the accurate phase,
|
650
|
+
#
|
651
|
+
# Should very bad cases be found in the future, only the degrees below
|
652
|
+
# should be increased, then this script rerun. Everything should
|
653
|
+
# compile OK with the newly generated trigo_accurate.h
|
654
|
+
|
655
|
+
printf("\n-------------------------------------------------------\n"):
|
656
|
+
printf("--------------------Accurate phase---------------------\n"):
|
657
|
+
printf("-------------------------------------------------------\n"):
|
658
|
+
|
659
|
+
xminSCS := 0:
|
660
|
+
xmaxSCS := Pi/4:
|
661
|
+
|
662
|
+
|
663
|
+
#-----------------Sine-----------------------
|
664
|
+
degreeSinSCS := 26:
|
665
|
+
Poly_P := convert(series(sin(sqrt(x))/(x^(3/2))-1/x, x=0, degreeSinSCS*4),polynom):
|
666
|
+
Poly_cheb := numapprox[chebpade](Poly_P, x=xminSCS..xmaxSCS^2, [degreeSinSCS/2-2,0]):
|
667
|
+
polySinSCS := poly_exact_SCS(expand(x + x^3 * subs(x=x^2, Poly_cheb))):
|
668
|
+
|
669
|
+
DEGREE_SIN_SCS := degree(polySinSCS):
|
670
|
+
maxepsApproxSinSCS:=numapprox[infnorm](1 - polySinSCS/sin(x), x=xminSCS..xmaxSCS):
|
671
|
+
printf("The sine polynomial for the second phase (degree %d) is accurate to %f bits on O..Pi/4\n", DEGREE_SIN_SCS, -log2(maxepsApproxSinSCS)):
|
672
|
+
# For the proof of the very bad cases, near zero
|
673
|
+
maxepsApproxSinSCSNearZero:=numapprox[infnorm](1 - polySinSCS/sin(x), x=0..2^(-17)):
|
674
|
+
printf(" ... and accurate to %f bits on O..2^(-17)\n", -log2(maxepsApproxSinSCSNearZero)):
|
675
|
+
|
676
|
+
|
677
|
+
#-----------------Cos-----------------------
|
678
|
+
degreeCosSCS := 28:
|
679
|
+
Poly_P := convert(series((cos(sqrt(x)) -1)/x, x=0, degreeCosSCS*4),polynom):
|
680
|
+
Poly_cheb := numapprox[chebpade](Poly_P, x=xminSCS..xmaxSCS^2, [degreeCosSCS/2-2,0]):
|
681
|
+
polyCosSCS := poly_exact_SCS(expand(1 + x^2 * subs(x=x^2, Poly_cheb))):
|
682
|
+
DEGREE_COS_SCS := degree(polyCosSCS):
|
683
|
+
|
684
|
+
maxepsApproxCosSCS:=numapprox[infnorm](1 - polyCosSCS/cos(x), x=xminSCS..xmaxSCS):
|
685
|
+
printf("The cos polynomial for the second phase (degree %d) is accurate to %f bits on O..Pi/4\n", degree(polyCosSCS), -log2(maxepsApproxCosSCS)):
|
686
|
+
# For the proof of the very bad cases, near zero
|
687
|
+
maxepsApproxCosSCSNearZero:=numapprox[infnorm](1 - polyCosSCS/cos(x), x=0..2^(-18)):
|
688
|
+
printf(" ... and accurate to %f bits on O..2^(-18)\n", -log2(maxepsApproxCosSCSNearZero)):
|
689
|
+
|
690
|
+
|
691
|
+
|
692
|
+
#-----------------Tan-----------------------
|
693
|
+
degreeTanSCS := 70:
|
694
|
+
Poly_P := convert(series(tan(sqrt(x))/(x^(3/2))-1/x, x=0, degreeTanSCS*4),polynom):
|
695
|
+
Poly_cheb := numapprox[chebpade](Poly_P, x=xminSCS..xmaxSCS^2, [degreeTanSCS/2-2,0]):
|
696
|
+
polyTanSCS := poly_exact_SCS(expand(x + x^3 * subs(x=x^2, Poly_cheb))):
|
697
|
+
|
698
|
+
DEGREE_TAN_SCS := degree(polyTanSCS):
|
699
|
+
maxepsApproxTanSCS:=numapprox[infnorm](1 - polyTanSCS/tan(x), x=xminSCS..xmaxSCS):
|
700
|
+
printf("The tan polynomial for the second phase (degree %d) is accurate to %f bits on O..Pi/4\n", DEGREE_TAN_SCS, -log2(maxepsApproxTanSCS)):
|
701
|
+
# For the proof of the very bad cases, near zero
|
702
|
+
maxepsApproxTanSCSNearZero:=numapprox[infnorm](1 - polyTanSCS/tan(x), x=0..2^(-17)):
|
703
|
+
printf(" ... and accurate to %f bits on O..2^(-17)\n", -log2(maxepsApproxTanSCSNearZero)):
|
704
|
+
|
705
|
+
|
706
|
+
filename:="TEMPTRIG/trigo_accurate.h":
|
707
|
+
fd:=fopen(filename, WRITE, TEXT):
|
708
|
+
fprintf(fd, "/*File generated by maple/trigo.mpl*/\n\n"):
|
709
|
+
fprintf(fd, "#include \"crlibm.h\"\n#include \"crlibm_private.h\"\n\n"):
|
710
|
+
fprintf(fd, "#define DEGREE_SIN_SCS %d \n", DEGREE_SIN_SCS):
|
711
|
+
fprintf(fd, "#define DEGREE_COS_SCS %d \n", DEGREE_COS_SCS):
|
712
|
+
fprintf(fd, "#define DEGREE_TAN_SCS %d \n", DEGREE_TAN_SCS):
|
713
|
+
fprintf(fd, "\n"):
|
714
|
+
fprintf(fd, "#define sin_scs_poly_ptr (scs_ptr)&sin_scs_poly \n"):
|
715
|
+
fprintf(fd, "#define cos_scs_poly_ptr (scs_ptr)&cos_scs_poly \n"):
|
716
|
+
fprintf(fd, "#define tan_scs_poly_ptr (scs_ptr)&tan_scs_poly \n"):
|
717
|
+
fprintf(fd, "\n"):
|
718
|
+
Write_SCS_poly(fd, sin_scs_poly, polySinSCS):
|
719
|
+
fprintf(fd, "\n\n"):
|
720
|
+
Write_SCS_poly(fd, cos_scs_poly, polyCosSCS):
|
721
|
+
fprintf(fd, "\n\n"):
|
722
|
+
Write_SCS_poly(fd, tan_scs_poly, polyTanSCS):
|
723
|
+
fprintf(fd, "\n"):
|
724
|
+
fclose(fd):
|
725
|
+
|
726
|
+
printf("\n\n************ DONE TEMPTRIG/trigo_accurate.h ************\n Copy it to the crlibm source directory.\n\n");
|
727
|
+
|
728
|
+
|