xref: /llvm-project/libclc/generic/lib/math/clc_rootn.cl (revision 78b5bb702fe97fe85f66d72598d0dfa7c49fe001)
1/*
2 * Copyright (c) 2014 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 <clc/clc.h>
24#include <clc/clcmacro.h>
25#include <clc/math/clc_fabs.h>
26#include <clc/math/clc_mad.h>
27#include <clc/math/clc_subnormal_config.h>
28#include <clc/math/math.h>
29#include <clc/math/tables.h>
30
31// compute pow using log and exp
32// x^y = exp(y * log(x))
33//
34// we take care not to lose precision in the intermediate steps
35//
36// When computing log, calculate it in splits,
37//
38// r = f * (p_invead + p_inv_tail)
39// r = rh + rt
40//
41// calculate log polynomial using r, in end addition, do
42// poly = poly + ((rh-r) + rt)
43//
44// lth = -r
45// ltt = ((xexp * log2_t) - poly) + logT
46// lt = lth + ltt
47//
48// lh = (xexp * log2_h) + logH
49// l = lh + lt
50//
51// Calculate final log answer as gh and gt,
52// gh = l & higher-half bits
53// gt = (((ltt - (lt - lth)) + ((lh - l) + lt)) + (l - gh))
54//
55// yh = y & higher-half bits
56// yt = y - yh
57//
58// Before entering computation of exp,
59// vs = ((yt*gt + yt*gh) + yh*gt)
60// v = vs + yh*gh
61// vt = ((yh*gh - v) + vs)
62//
63// In calculation of exp, add vt to r that is used for poly
64// At the end of exp, do
65// ((((expT * poly) + expT) + expH*poly) + expH)
66
67_CLC_DEF _CLC_OVERLOAD float __clc_rootn(float x, int ny) {
68  float y = MATH_RECIP((float)ny);
69
70  int ix = as_int(x);
71  int ax = ix & EXSIGNBIT_SP32;
72  int xpos = ix == ax;
73
74  int iy = as_int(y);
75  int ay = iy & EXSIGNBIT_SP32;
76  int ypos = iy == ay;
77
78  // Extra precise log calculation
79  // First handle case that x is close to 1
80  float r = 1.0f - as_float(ax);
81  int near1 = __clc_fabs(r) < 0x1.0p-4f;
82  float r2 = r * r;
83
84  // Coefficients are just 1/3, 1/4, 1/5 and 1/6
85  float poly = __clc_mad(
86      r,
87      __clc_mad(r,
88                __clc_mad(r, __clc_mad(r, 0x1.24924ap-3f, 0x1.555556p-3f),
89                          0x1.99999ap-3f),
90                0x1.000000p-2f),
91      0x1.555556p-2f);
92
93  poly *= r2 * r;
94
95  float lth_near1 = -r2 * 0.5f;
96  float ltt_near1 = -poly;
97  float lt_near1 = lth_near1 + ltt_near1;
98  float lh_near1 = -r;
99  float l_near1 = lh_near1 + lt_near1;
100
101  // Computations for x not near 1
102  int m = (int)(ax >> EXPSHIFTBITS_SP32) - EXPBIAS_SP32;
103  float mf = (float)m;
104  int ixs = as_int(as_float(ax | 0x3f800000) - 1.0f);
105  float mfs = (float)((ixs >> EXPSHIFTBITS_SP32) - 253);
106  int c = m == -127;
107  int ixn = c ? ixs : ax;
108  float mfn = c ? mfs : mf;
109
110  int indx = (ixn & 0x007f0000) + ((ixn & 0x00008000) << 1);
111
112  // F - Y
113  float f = as_float(0x3f000000 | indx) -
114            as_float(0x3f000000 | (ixn & MANTBITS_SP32));
115
116  indx = indx >> 16;
117  float2 tv = USE_TABLE(log_inv_tbl_ep, indx);
118  float rh = f * tv.s0;
119  float rt = f * tv.s1;
120  r = rh + rt;
121
122  poly = __clc_mad(r, __clc_mad(r, 0x1.0p-2f, 0x1.555556p-2f), 0x1.0p-1f) *
123         (r * r);
124  poly += (rh - r) + rt;
125
126  const float LOG2_HEAD = 0x1.62e000p-1f;  // 0.693115234
127  const float LOG2_TAIL = 0x1.0bfbe8p-15f; // 0.0000319461833
128  tv = USE_TABLE(loge_tbl, indx);
129  float lth = -r;
130  float ltt = __clc_mad(mfn, LOG2_TAIL, -poly) + tv.s1;
131  float lt = lth + ltt;
132  float lh = __clc_mad(mfn, LOG2_HEAD, tv.s0);
133  float l = lh + lt;
134
135  // Select near 1 or not
136  lth = near1 ? lth_near1 : lth;
137  ltt = near1 ? ltt_near1 : ltt;
138  lt = near1 ? lt_near1 : lt;
139  lh = near1 ? lh_near1 : lh;
140  l = near1 ? l_near1 : l;
141
142  float gh = as_float(as_int(l) & 0xfffff000);
143  float gt = ((ltt - (lt - lth)) + ((lh - l) + lt)) + (l - gh);
144
145  float yh = as_float(iy & 0xfffff000);
146
147  float fny = (float)ny;
148  float fnyh = as_float(as_int(fny) & 0xfffff000);
149  float fnyt = (float)(ny - (int)fnyh);
150  float yt = MATH_DIVIDE(__clc_mad(-fnyt, yh, __clc_mad(-fnyh, yh, 1.0f)), fny);
151
152  float ylogx_s = __clc_mad(gt, yh, __clc_mad(gh, yt, yt * gt));
153  float ylogx = __clc_mad(yh, gh, ylogx_s);
154  float ylogx_t = __clc_mad(yh, gh, -ylogx) + ylogx_s;
155
156  // Extra precise exp of ylogx
157  const float R_64_BY_LOG2 = 0x1.715476p+6f; // 64/log2 : 92.332482616893657
158  int n = convert_int(ylogx * R_64_BY_LOG2);
159  float nf = (float)n;
160
161  int j = n & 0x3f;
162  m = n >> 6;
163  int m2 = m << EXPSHIFTBITS_SP32;
164
165  // log2/64 lead: 0.0108032227
166  const float R_LOG2_BY_64_LD = 0x1.620000p-7f;
167  // log2/64 tail: 0.0000272020388
168  const float R_LOG2_BY_64_TL = 0x1.c85fdep-16f;
169  r = __clc_mad(nf, -R_LOG2_BY_64_TL, __clc_mad(nf, -R_LOG2_BY_64_LD, ylogx)) +
170      ylogx_t;
171
172  // Truncated Taylor series for e^r
173  poly = __clc_mad(__clc_mad(__clc_mad(r, 0x1.555556p-5f, 0x1.555556p-3f), r,
174                             0x1.000000p-1f),
175                   r * r, r);
176
177  tv = USE_TABLE(exp_tbl_ep, j);
178
179  float expylogx =
180      __clc_mad(tv.s0, poly, __clc_mad(tv.s1, poly, tv.s1)) + tv.s0;
181  float sexpylogx = __clc_fp32_subnormals_supported()
182                        ? expylogx * as_float(0x1 << (m + 149))
183                        : 0.0f;
184
185  float texpylogx = as_float(as_int(expylogx) + m2);
186  expylogx = m < -125 ? sexpylogx : texpylogx;
187
188  // Result is +-Inf if (ylogx + ylogx_t) > 128*log2
189  expylogx = ((ylogx > 0x1.62e430p+6f) |
190              (ylogx == 0x1.62e430p+6f & ylogx_t > -0x1.05c610p-22f))
191                 ? as_float(PINFBITPATT_SP32)
192                 : expylogx;
193
194  // Result is 0 if ylogx < -149*log2
195  expylogx = ylogx < -0x1.9d1da0p+6f ? 0.0f : expylogx;
196
197  // Classify y:
198  //   inty = 0 means not an integer.
199  //   inty = 1 means odd integer.
200  //   inty = 2 means even integer.
201
202  int inty = 2 - (ny & 1);
203
204  float signval = as_float((as_uint(expylogx) ^ SIGNBIT_SP32));
205  expylogx = ((inty == 1) & !xpos) ? signval : expylogx;
206  int ret = as_int(expylogx);
207
208  // Corner case handling
209  ret = (!xpos & (inty == 2)) ? QNANBITPATT_SP32 : ret;
210  int xinf = xpos ? PINFBITPATT_SP32 : NINFBITPATT_SP32;
211  ret = ((ax == 0) & !ypos & (inty == 1)) ? xinf : ret;
212  ret = ((ax == 0) & !ypos & (inty == 2)) ? PINFBITPATT_SP32 : ret;
213  ret = ((ax == 0) & ypos & (inty == 2)) ? 0 : ret;
214  int xzero = xpos ? 0 : 0x80000000;
215  ret = ((ax == 0) & ypos & (inty == 1)) ? xzero : ret;
216  ret =
217      ((ix == NINFBITPATT_SP32) & ypos & (inty == 1)) ? NINFBITPATT_SP32 : ret;
218  ret = ((ix == NINFBITPATT_SP32) & !ypos & (inty == 1)) ? 0x80000000 : ret;
219  ret = ((ix == PINFBITPATT_SP32) & !ypos) ? 0 : ret;
220  ret = ((ix == PINFBITPATT_SP32) & ypos) ? PINFBITPATT_SP32 : ret;
221  ret = ax > PINFBITPATT_SP32 ? ix : ret;
222  ret = ny == 0 ? QNANBITPATT_SP32 : ret;
223
224  return as_float(ret);
225}
226_CLC_BINARY_VECTORIZE(_CLC_DEF _CLC_OVERLOAD, float, __clc_rootn, float, int)
227
228#ifdef cl_khr_fp64
229_CLC_DEF _CLC_OVERLOAD double __clc_rootn(double x, int ny) {
230  const double real_log2_tail = 5.76999904754328540596e-08;
231  const double real_log2_lead = 6.93147122859954833984e-01;
232
233  double dny = (double)ny;
234  double y = 1.0 / dny;
235
236  long ux = as_long(x);
237  long ax = ux & (~SIGNBIT_DP64);
238  int xpos = ax == ux;
239
240  long uy = as_long(y);
241  long ay = uy & (~SIGNBIT_DP64);
242  int ypos = ay == uy;
243
244  // Extended precision log
245  double v, vt;
246  {
247    int exp = (int)(ax >> 52) - 1023;
248    int mask_exp_1023 = exp == -1023;
249    double xexp = (double)exp;
250    long mantissa = ax & 0x000FFFFFFFFFFFFFL;
251
252    long temp_ux = as_long(as_double(0x3ff0000000000000L | mantissa) - 1.0);
253    exp = ((temp_ux & 0x7FF0000000000000L) >> 52) - 2045;
254    double xexp1 = (double)exp;
255    long mantissa1 = temp_ux & 0x000FFFFFFFFFFFFFL;
256
257    xexp = mask_exp_1023 ? xexp1 : xexp;
258    mantissa = mask_exp_1023 ? mantissa1 : mantissa;
259
260    long rax = (mantissa & 0x000ff00000000000) +
261               ((mantissa & 0x0000080000000000) << 1);
262    int index = rax >> 44;
263
264    double F = as_double(rax | 0x3FE0000000000000L);
265    double Y = as_double(mantissa | 0x3FE0000000000000L);
266    double f = F - Y;
267    double2 tv = USE_TABLE(log_f_inv_tbl, index);
268    double log_h = tv.s0;
269    double log_t = tv.s1;
270    double f_inv = (log_h + log_t) * f;
271    double r1 = as_double(as_long(f_inv) & 0xfffffffff8000000L);
272    double r2 = fma(-F, r1, f) * (log_h + log_t);
273    double r = r1 + r2;
274
275    double poly = fma(
276        r, fma(r, fma(r, fma(r, 1.0 / 7.0, 1.0 / 6.0), 1.0 / 5.0), 1.0 / 4.0),
277        1.0 / 3.0);
278    poly = poly * r * r * r;
279
280    double hr1r1 = 0.5 * r1 * r1;
281    double poly0h = r1 + hr1r1;
282    double poly0t = r1 - poly0h + hr1r1;
283    poly = fma(r1, r2, fma(0.5 * r2, r2, poly)) + r2 + poly0t;
284
285    tv = USE_TABLE(powlog_tbl, index);
286    log_h = tv.s0;
287    log_t = tv.s1;
288
289    double resT_t = fma(xexp, real_log2_tail, +log_t) - poly;
290    double resT = resT_t - poly0h;
291    double resH = fma(xexp, real_log2_lead, log_h);
292    double resT_h = poly0h;
293
294    double H = resT + resH;
295    double H_h = as_double(as_long(H) & 0xfffffffff8000000L);
296    double T = (resH - H + resT) + (resT_t - (resT + resT_h)) + (H - H_h);
297    H = H_h;
298
299    double y_head = as_double(uy & 0xfffffffff8000000L);
300    double y_tail = y - y_head;
301
302    double fnyh = as_double(as_long(dny) & 0xfffffffffff00000);
303    double fnyt = (double)(ny - (int)fnyh);
304    y_tail = fma(-fnyt, y_head, fma(-fnyh, y_head, 1.0)) / dny;
305
306    double temp = fma(y_tail, H, fma(y_head, T, y_tail * T));
307    v = fma(y_head, H, temp);
308    vt = fma(y_head, H, -v) + temp;
309  }
310
311  // Now calculate exp of (v,vt)
312
313  double expv;
314  {
315    const double max_exp_arg = 709.782712893384;
316    const double min_exp_arg = -745.1332191019411;
317    const double sixtyfour_by_lnof2 = 92.33248261689366;
318    const double lnof2_by_64_head = 0.010830424260348081;
319    const double lnof2_by_64_tail = -4.359010638708991e-10;
320
321    double temp = v * sixtyfour_by_lnof2;
322    int n = (int)temp;
323    double dn = (double)n;
324    int j = n & 0x0000003f;
325    int m = n >> 6;
326
327    double2 tv = USE_TABLE(two_to_jby64_ep_tbl, j);
328    double f1 = tv.s0;
329    double f2 = tv.s1;
330    double f = f1 + f2;
331
332    double r1 = fma(dn, -lnof2_by_64_head, v);
333    double r2 = dn * lnof2_by_64_tail;
334    double r = (r1 + r2) + vt;
335
336    double q = fma(
337        r,
338        fma(r,
339            fma(r,
340                fma(r, 1.38889490863777199667e-03, 8.33336798434219616221e-03),
341                4.16666666662260795726e-02),
342            1.66666666665260878863e-01),
343        5.00000000000000008883e-01);
344    q = fma(r * r, q, r);
345
346    expv = fma(f, q, f2) + f1;
347    expv = ldexp(expv, m);
348
349    expv = v > max_exp_arg ? as_double(0x7FF0000000000000L) : expv;
350    expv = v < min_exp_arg ? 0.0 : expv;
351  }
352
353  // See whether y is an integer.
354  // inty = 0 means not an integer.
355  // inty = 1 means odd integer.
356  // inty = 2 means even integer.
357
358  int inty = 2 - (ny & 1);
359
360  expv *= ((inty == 1) & !xpos) ? -1.0 : 1.0;
361
362  long ret = as_long(expv);
363
364  // Now all the edge cases
365  ret = (!xpos & (inty == 2)) ? QNANBITPATT_DP64 : ret;
366  long xinf = xpos ? PINFBITPATT_DP64 : NINFBITPATT_DP64;
367  ret = ((ax == 0L) & !ypos & (inty == 1)) ? xinf : ret;
368  ret = ((ax == 0L) & !ypos & (inty == 2)) ? PINFBITPATT_DP64 : ret;
369  ret = ((ax == 0L) & ypos & (inty == 2)) ? 0L : ret;
370  long xzero = xpos ? 0L : 0x8000000000000000L;
371  ret = ((ax == 0L) & ypos & (inty == 1)) ? xzero : ret;
372  ret =
373      ((ux == NINFBITPATT_DP64) & ypos & (inty == 1)) ? NINFBITPATT_DP64 : ret;
374  ret = ((ux == NINFBITPATT_DP64) & !ypos & (inty == 1)) ? 0x8000000000000000L
375                                                         : ret;
376  ret = ((ux == PINFBITPATT_DP64) & !ypos) ? 0L : ret;
377  ret = ((ux == PINFBITPATT_DP64) & ypos) ? PINFBITPATT_DP64 : ret;
378  ret = ax > PINFBITPATT_DP64 ? ux : ret;
379  ret = ny == 0 ? QNANBITPATT_DP64 : ret;
380  return as_double(ret);
381}
382_CLC_BINARY_VECTORIZE(_CLC_DEF _CLC_OVERLOAD, double, __clc_rootn, double, int)
383#endif
384
385#ifdef cl_khr_fp16
386
387#pragma OPENCL EXTENSION cl_khr_fp16 : enable
388
389_CLC_OVERLOAD _CLC_DEF half __clc_rootn(half x, int y) {
390    return (half)__clc_rootn((float)x, y);
391}
392
393_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, half, __clc_rootn, half, int);
394
395#endif
396