• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2014,2015 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a copy
5  * of this software and associated documentation files (the "Software"), to deal
6  * in the Software without restriction, including without limitation the rights
7  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8  * copies of the Software, and to permit persons to whom the Software is
9  * furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
20  * THE SOFTWARE.
21  */
22 
23 #include "math.h"
24 
25 /*
26    Algorithm:
27 
28    Based on:
29    Ping-Tak Peter Tang
30    "Table-driven implementation of the logarithm function in IEEE
31    floating-point arithmetic"
32    ACM Transactions on Mathematical Software (TOMS)
33    Volume 16, Issue 4 (December 1990)
34 
35 
36    x very close to 1.0 is handled differently, for x everywhere else
37    a brief explanation is given below
38 
39    x = (2^m)*A
40    x = (2^m)*(G+g) with (1 <= G < 2) and (g <= 2^(-8))
41    x = (2^m)*2*(G/2+g/2)
42    x = (2^m)*2*(F+f) with (0.5 <= F < 1) and (f <= 2^(-9))
43 
44    Y = (2^(-1))*(2^(-m))*(2^m)*A
45    Now, range of Y is: 0.5 <= Y < 1
46 
47    F = 0x80 + (first 7 mantissa bits) + (8th mantissa bit)
48    Now, range of F is: 128 <= F <= 256
49    F = F / 256
50    Now, range of F is: 0.5 <= F <= 1
51 
52    f = -(Y-F), with (f <= 2^(-9))
53 
54    log(x) = m*log(2) + log(2) + log(F-f)
55    log(x) = m*log(2) + log(2) + log(F) + log(1-(f/F))
56    log(x) = m*log(2) + log(2*F) + log(1-r)
57 
58    r = (f/F), with (r <= 2^(-8))
59    r = f*(1/F) with (1/F) precomputed to avoid division
60 
61    log(x) = m*log(2) + log(G) - poly
62 
63    log(G) is precomputed
64    poly = (r + (r^2)/2 + (r^3)/3 + (r^4)/4) + (r^5)/5))
65 
66    log(2) and log(G) need to be maintained in extra precision
67    to avoid losing precision in the calculations
68 
69 
70    For x close to 1.0, we employ the following technique to
71    ensure faster convergence.
72 
73    log(x) = log((1+s)/(1-s)) = 2*s + (2/3)*s^3 + (2/5)*s^5 + (2/7)*s^7
74    x = ((1+s)/(1-s))
75    x = 1 + r
76    s = r/(2+r)
77 
78 */
79 
80 _CLC_OVERLOAD _CLC_DEF float
81 #if defined(COMPILING_LOG2)
log2(float x)82 log2(float x)
83 #elif defined(COMPILING_LOG10)
84 log10(float x)
85 #else
86 log(float x)
87 #endif
88 {
89 
90 #if defined(COMPILING_LOG2)
91     const float LOG2E = 0x1.715476p+0f;      // 1.4426950408889634
92     const float LOG2E_HEAD = 0x1.700000p+0f; // 1.4375
93     const float LOG2E_TAIL = 0x1.547652p-8f; // 0.00519504072
94 #elif defined(COMPILING_LOG10)
95     const float LOG10E = 0x1.bcb7b2p-2f;        // 0.43429448190325182
96     const float LOG10E_HEAD = 0x1.bc0000p-2f;   // 0.43359375
97     const float LOG10E_TAIL = 0x1.6f62a4p-11f;  // 0.0007007319
98     const float LOG10_2_HEAD = 0x1.340000p-2f;  // 0.30078125
99     const float LOG10_2_TAIL = 0x1.04d426p-12f; // 0.000248745637
100 #else
101     const float LOG2_HEAD = 0x1.62e000p-1f;  // 0.693115234
102     const float LOG2_TAIL = 0x1.0bfbe8p-15f; // 0.0000319461833
103 #endif
104 
105     uint xi = as_uint(x);
106     uint ax = xi & EXSIGNBIT_SP32;
107 
108     // Calculations for |x-1| < 2^-4
109     float r = x - 1.0f;
110     int near1 = fabs(r) < 0x1.0p-4f;
111     float u2 = MATH_DIVIDE(r, 2.0f + r);
112     float corr = u2 * r;
113     float u = u2 + u2;
114     float v = u * u;
115     float znear1, z1, z2;
116 
117     // 2/(5 * 2^5), 2/(3 * 2^3)
118     z2 = mad(u, mad(v, 0x1.99999ap-7f, 0x1.555556p-4f)*v, -corr);
119 
120 #if defined(COMPILING_LOG2)
121     z1 = as_float(as_int(r) & 0xffff0000);
122     z2 = z2 + (r - z1);
123     znear1 = mad(z1, LOG2E_HEAD, mad(z2, LOG2E_HEAD, mad(z1, LOG2E_TAIL, z2*LOG2E_TAIL)));
124 #elif defined(COMPILING_LOG10)
125     z1 = as_float(as_int(r) & 0xffff0000);
126     z2 = z2 + (r - z1);
127     znear1 = mad(z1, LOG10E_HEAD, mad(z2, LOG10E_HEAD, mad(z1, LOG10E_TAIL, z2*LOG10E_TAIL)));
128 #else
129     znear1 = z2 + r;
130 #endif
131 
132     // Calculations for x not near 1
133     int m = (int)(xi >> EXPSHIFTBITS_SP32) - EXPBIAS_SP32;
134 
135     // Normalize subnormal
136     uint xis = as_uint(as_float(xi | 0x3f800000) - 1.0f);
137     int ms = (int)(xis >> EXPSHIFTBITS_SP32) - 253;
138     int c = m == -127;
139     m = c ? ms : m;
140     uint xin = c ? xis : xi;
141 
142     float mf = (float)m;
143     uint indx = (xin & 0x007f0000) + ((xin & 0x00008000) << 1);
144 
145     // F - Y
146     float f = as_float(0x3f000000 | indx) - as_float(0x3f000000 | (xin & MANTBITS_SP32));
147 
148     indx = indx >> 16;
149     r = f * USE_TABLE(log_inv_tbl, indx);
150 
151     // 1/3,  1/2
152     float poly = mad(mad(r, 0x1.555556p-2f, 0.5f), r*r, r);
153 
154 #if defined(COMPILING_LOG2)
155     float2 tv = USE_TABLE(log2_tbl, indx);
156     z1 = tv.s0 + mf;
157     z2 = mad(poly, -LOG2E, tv.s1);
158 #elif defined(COMPILING_LOG10)
159     float2 tv = USE_TABLE(log10_tbl, indx);
160     z1 = mad(mf, LOG10_2_HEAD, tv.s0);
161     z2 = mad(poly, -LOG10E, mf*LOG10_2_TAIL) + tv.s1;
162 #else
163     float2 tv = USE_TABLE(log_tbl, indx);
164     z1 = mad(mf, LOG2_HEAD, tv.s0);
165     z2 = mad(mf, LOG2_TAIL, -poly) + tv.s1;
166 #endif
167 
168     float z = z1 + z2;
169     z = near1 ? znear1 : z;
170 
171     // Corner cases
172     z = ax >= PINFBITPATT_SP32 ? x : z;
173     z = xi != ax ? as_float(QNANBITPATT_SP32) : z;
174     z = ax == 0 ? as_float(NINFBITPATT_SP32) : z;
175 
176     return z;
177 }
178 
179 #ifdef cl_khr_fp64
180 
181 _CLC_OVERLOAD _CLC_DEF double
182 #if defined(COMPILING_LOG2)
log2(double x)183 log2(double x)
184 #elif defined(COMPILING_LOG10)
185 log10(double x)
186 #else
187 log(double x)
188 #endif
189 {
190 
191 #ifndef COMPILING_LOG2
192     // log2_lead and log2_tail sum to an extra-precise version of ln(2)
193     const double log2_lead = 6.93147122859954833984e-01; /* 0x3fe62e42e0000000 */
194     const double log2_tail = 5.76999904754328540596e-08; /* 0x3e6efa39ef35793c */
195 #endif
196 
197 #if defined(COMPILING_LOG10)
198     // log10e_lead and log10e_tail sum to an extra-precision version of log10(e) (19 bits in lead)
199     const double log10e_lead = 4.34293746948242187500e-01;  /* 0x3fdbcb7800000000 */
200     const double log10e_tail = 7.3495500964015109100644e-7; /* 0x3ea8a93728719535 */
201 #elif defined(COMPILING_LOG2)
202     // log2e_lead and log2e_tail sum to an extra-precision version of log2(e) (19 bits in lead)
203     const double log2e_lead = 1.44269180297851562500E+00; /* 0x3FF7154400000000 */
204     const double log2e_tail = 3.23791044778235969970E-06; /* 0x3ECB295C17F0BBBE */
205 #endif
206 
207     // log_thresh1 = 9.39412117004394531250e-1 = 0x3fee0faa00000000
208     // log_thresh2 = 1.06449508666992187500 = 0x3ff1082c00000000
209     const double log_thresh1 = 0x1.e0faap-1;
210     const double log_thresh2 = 0x1.1082cp+0;
211 
212     int is_near = x >= log_thresh1 & x <= log_thresh2;
213 
214     // Near 1 code
215     double r = x - 1.0;
216     double u = r / (2.0 + r);
217     double correction = r * u;
218     u = u + u;
219     double v = u * u;
220     double r1 = r;
221 
222     const double ca_1 = 8.33333333333317923934e-02; /* 0x3fb55555555554e6 */
223     const double ca_2 = 1.25000000037717509602e-02; /* 0x3f89999999bac6d4 */
224     const double ca_3 = 2.23213998791944806202e-03; /* 0x3f62492307f1519f */
225     const double ca_4 = 4.34887777707614552256e-04; /* 0x3f3c8034c85dfff0 */
226 
227     double r2 = fma(u*v, fma(v, fma(v, fma(v, ca_4, ca_3), ca_2), ca_1), -correction);
228 
229 #if defined(COMPILING_LOG10)
230     r = r1;
231     r1 = as_double(as_ulong(r1) & 0xffffffff00000000);
232     r2 = r2 + (r - r1);
233     double ret_near = fma(log10e_lead, r1, fma(log10e_lead, r2, fma(log10e_tail, r1, log10e_tail * r2)));
234 #elif defined(COMPILING_LOG2)
235     r = r1;
236     r1 = as_double(as_ulong(r1) & 0xffffffff00000000);
237     r2 = r2 + (r - r1);
238     double ret_near = fma(log2e_lead, r1, fma(log2e_lead, r2, fma(log2e_tail, r1, log2e_tail*r2)));
239 #else
240     double ret_near = r1 + r2;
241 #endif
242 
243     // This is the far from 1 code
244 
245     // Deal with subnormal
246     ulong ux = as_ulong(x);
247     ulong uxs = as_ulong(as_double(0x03d0000000000000UL | ux) - 0x1.0p-962);
248     int c = ux < IMPBIT_DP64;
249     ux = c ? uxs : ux;
250     int expadjust = c ? 60 : 0;
251 
252     int xexp = ((as_int2(ux).hi >> 20) & 0x7ff) - EXPBIAS_DP64 - expadjust;
253     double f = as_double(HALFEXPBITS_DP64 | (ux & MANTBITS_DP64));
254     int index = as_int2(ux).hi >> 13;
255     index = ((0x80 | (index & 0x7e)) >> 1) + (index & 0x1);
256 
257     double2 tv = USE_TABLE(ln_tbl, index - 64);
258     double z1 = tv.s0;
259     double q = tv.s1;
260 
261     double f1 = index * 0x1.0p-7;
262     double f2 = f - f1;
263     u = f2 / fma(f2, 0.5, f1);
264     v = u * u;
265 
266     const double cb_1 = 8.33333333333333593622e-02; /* 0x3fb5555555555557 */
267     const double cb_2 = 1.24999999978138668903e-02; /* 0x3f89999999865ede */
268     const double cb_3 = 2.23219810758559851206e-03; /* 0x3f6249423bd94741 */
269 
270     double poly = v * fma(v, fma(v, cb_3, cb_2), cb_1);
271     double z2 = q + fma(u, poly, u);
272 
273     double dxexp = (double)xexp;
274 #if defined (COMPILING_LOG10)
275     // Add xexp * log(2) to z1,z2 to get log(x)
276     r1 = fma(dxexp, log2_lead, z1);
277     r2 = fma(dxexp, log2_tail, z2);
278     double ret_far = fma(log10e_lead, r1, fma(log10e_lead, r2, fma(log10e_tail, r1, log10e_tail*r2)));
279 #elif defined(COMPILING_LOG2)
280     r1 = fma(log2e_lead, z1, dxexp);
281     r2 = fma(log2e_lead, z2, fma(log2e_tail, z1, log2e_tail*z2));
282     double ret_far = r1 + r2;
283 #else
284     r1 = fma(dxexp, log2_lead, z1);
285     r2 = fma(dxexp, log2_tail, z2);
286     double ret_far = r1 + r2;
287 #endif
288 
289     double ret = is_near ? ret_near : ret_far;
290 
291     ret = isinf(x) ? as_double(PINFBITPATT_DP64) : ret;
292     ret = isnan(x) | (x < 0.0) ? as_double(QNANBITPATT_DP64) : ret;
293     ret = x == 0.0 ? as_double(NINFBITPATT_DP64) : ret;
294     return ret;
295 }
296 
297 #endif // cl_khr_fp64
298