1 /* mpn_mu_div_qr, mpn_preinv_mu_div_qr.
2 
3    Compute Q = floor(N / D) and R = N-QD.  N is nn limbs and D is dn limbs and
4    must be normalized, and Q must be nn-dn limbs.  The requirement that Q is
5    nn-dn limbs (and not nn-dn+1 limbs) was put in place in order to allow us to
6    let N be unmodified during the operation.
7 
8    Contributed to the GNU project by Torbjorn Granlund.
9 
10    THE FUNCTIONS IN THIS FILE ARE INTERNAL WITH MUTABLE INTERFACES.  IT IS ONLY
11    SAFE TO REACH THEM THROUGH DOCUMENTED INTERFACES.  IN FACT, IT IS ALMOST
12    GUARANTEED THAT THEY WILL CHANGE OR DISAPPEAR IN A FUTURE GMP RELEASE.
13 
14 Copyright 2005-2007, 2009, 2010 Free Software Foundation, Inc.
15 
16 This file is part of the GNU MP Library.
17 
18 The GNU MP Library is free software; you can redistribute it and/or modify
19 it under the terms of either:
20 
21   * the GNU Lesser General Public License as published by the Free
22     Software Foundation; either version 3 of the License, or (at your
23     option) any later version.
24 
25 or
26 
27   * the GNU General Public License as published by the Free Software
28     Foundation; either version 2 of the License, or (at your option) any
29     later version.
30 
31 or both in parallel, as here.
32 
33 The GNU MP Library is distributed in the hope that it will be useful, but
34 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
35 or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
36 for more details.
37 
38 You should have received copies of the GNU General Public License and the
39 GNU Lesser General Public License along with the GNU MP Library.  If not,
40 see https://www.gnu.org/licenses/.  */
41 
42 
43 /*
44    The idea of the algorithm used herein is to compute a smaller inverted value
45    than used in the standard Barrett algorithm, and thus save time in the
46    Newton iterations, and pay just a small price when using the inverted value
47    for developing quotient bits.  This algorithm was presented at ICMS 2006.
48 */
49 
50 /* CAUTION: This code and the code in mu_divappr_q.c should be edited in sync.
51 
52  Things to work on:
53 
54   * This isn't optimal when the quotient isn't needed, as it might take a lot
55     of space.  The computation is always needed, though, so there is no time to
56     save with special code.
57 
58   * The itch/scratch scheme isn't perhaps such a good idea as it once seemed,
59     demonstrated by the fact that the mpn_invertappr function's scratch needs
60     mean that we need to keep a large allocation long after it is needed.
61     Things are worse as mpn_mul_fft does not accept any scratch parameter,
62     which means we'll have a large memory hole while in mpn_mul_fft.  In
63     general, a peak scratch need in the beginning of a function isn't
64     well-handled by the itch/scratch scheme.
65 */
66 
67 #ifdef STAT
68 #undef STAT
69 #define STAT(x) x
70 #else
71 #define STAT(x)
72 #endif
73 
74 #include <stdlib.h>		/* for NULL */
75 #include "gmp-impl.h"
76 
77 
78 /* FIXME: The MU_DIV_QR_SKEW_THRESHOLD was not analysed properly.  It gives a
79    speedup according to old measurements, but does the decision mechanism
80    really make sense?  It seem like the quotient between dn and qn might be
81    what we really should be checking.  */
82 #ifndef MU_DIV_QR_SKEW_THRESHOLD
83 #define MU_DIV_QR_SKEW_THRESHOLD 100
84 #endif
85 
86 #ifdef CHECK				/* FIXME: Enable in minithres */
87 #undef  MU_DIV_QR_SKEW_THRESHOLD
88 #define MU_DIV_QR_SKEW_THRESHOLD 1
89 #endif
90 
91 
92 static mp_limb_t mpn_mu_div_qr2 (mp_ptr, mp_ptr, mp_srcptr, mp_size_t, mp_srcptr, mp_size_t, mp_ptr);
93 static mp_size_t mpn_mu_div_qr_choose_in (mp_size_t, mp_size_t, int);
94 
95 
96 mp_limb_t
mpn_mu_div_qr(mp_ptr qp,mp_ptr rp,mp_srcptr np,mp_size_t nn,mp_srcptr dp,mp_size_t dn,mp_ptr scratch)97 mpn_mu_div_qr (mp_ptr qp,
98 	       mp_ptr rp,
99 	       mp_srcptr np,
100 	       mp_size_t nn,
101 	       mp_srcptr dp,
102 	       mp_size_t dn,
103 	       mp_ptr scratch)
104 {
105   mp_size_t qn;
106   mp_limb_t cy, qh;
107 
108   qn = nn - dn;
109   if (qn + MU_DIV_QR_SKEW_THRESHOLD < dn)
110     {
111       /* |______________|_ign_first__|   dividend			  nn
112 		|_______|_ign_first__|   divisor			  dn
113 
114 		|______|	     quotient (prel)			  qn
115 
116 		 |___________________|   quotient * ignored-divisor-part  dn-1
117       */
118 
119       /* Compute a preliminary quotient and a partial remainder by dividing the
120 	 most significant limbs of each operand.  */
121       qh = mpn_mu_div_qr2 (qp, rp + nn - (2 * qn + 1),
122 			   np + nn - (2 * qn + 1), 2 * qn + 1,
123 			   dp + dn - (qn + 1), qn + 1,
124 			   scratch);
125 
126       /* Multiply the quotient by the divisor limbs ignored above.  */
127       if (dn - (qn + 1) > qn)
128 	mpn_mul (scratch, dp, dn - (qn + 1), qp, qn);  /* prod is dn-1 limbs */
129       else
130 	mpn_mul (scratch, qp, qn, dp, dn - (qn + 1));  /* prod is dn-1 limbs */
131 
132       if (qh)
133 	cy = mpn_add_n (scratch + qn, scratch + qn, dp, dn - (qn + 1));
134       else
135 	cy = 0;
136       scratch[dn - 1] = cy;
137 
138       cy = mpn_sub_n (rp, np, scratch, nn - (2 * qn + 1));
139       cy = mpn_sub_nc (rp + nn - (2 * qn + 1),
140 		       rp + nn - (2 * qn + 1),
141 		       scratch + nn - (2 * qn + 1),
142 		       qn + 1, cy);
143       if (cy)
144 	{
145 	  qh -= mpn_sub_1 (qp, qp, qn, 1);
146 	  mpn_add_n (rp, rp, dp, dn);
147 	}
148     }
149   else
150     {
151       qh = mpn_mu_div_qr2 (qp, rp, np, nn, dp, dn, scratch);
152     }
153 
154   return qh;
155 }
156 
157 static mp_limb_t
mpn_mu_div_qr2(mp_ptr qp,mp_ptr rp,mp_srcptr np,mp_size_t nn,mp_srcptr dp,mp_size_t dn,mp_ptr scratch)158 mpn_mu_div_qr2 (mp_ptr qp,
159 		mp_ptr rp,
160 		mp_srcptr np,
161 		mp_size_t nn,
162 		mp_srcptr dp,
163 		mp_size_t dn,
164 		mp_ptr scratch)
165 {
166   mp_size_t qn, in;
167   mp_limb_t cy, qh;
168   mp_ptr ip, tp;
169 
170   ASSERT (dn > 1);
171 
172   qn = nn - dn;
173 
174   /* Compute the inverse size.  */
175   in = mpn_mu_div_qr_choose_in (qn, dn, 0);
176   ASSERT (in <= dn);
177 
178 #if 1
179   /* This alternative inverse computation method gets slightly more accurate
180      results.  FIXMEs: (1) Temp allocation needs not analysed (2) itch function
181      not adapted (3) mpn_invertappr scratch needs not met.  */
182   ip = scratch;
183   tp = scratch + in + 1;
184 
185   /* compute an approximate inverse on (in+1) limbs */
186   if (dn == in)
187     {
188       MPN_COPY (tp + 1, dp, in);
189       tp[0] = 1;
190       mpn_invertappr (ip, tp, in + 1, tp + in + 1);
191       MPN_COPY_INCR (ip, ip + 1, in);
192     }
193   else
194     {
195       cy = mpn_add_1 (tp, dp + dn - (in + 1), in + 1, 1);
196       if (UNLIKELY (cy != 0))
197 	MPN_ZERO (ip, in);
198       else
199 	{
200 	  mpn_invertappr (ip, tp, in + 1, tp + in + 1);
201 	  MPN_COPY_INCR (ip, ip + 1, in);
202 	}
203     }
204 #else
205   /* This older inverse computation method gets slightly worse results than the
206      one above.  */
207   ip = scratch;
208   tp = scratch + in;
209 
210   /* Compute inverse of D to in+1 limbs, then round to 'in' limbs.  Ideally the
211      inversion function should do this automatically.  */
212   if (dn == in)
213     {
214       tp[in + 1] = 0;
215       MPN_COPY (tp + in + 2, dp, in);
216       mpn_invertappr (tp, tp + in + 1, in + 1, NULL);
217     }
218   else
219     {
220       mpn_invertappr (tp, dp + dn - (in + 1), in + 1, NULL);
221     }
222   cy = mpn_sub_1 (tp, tp, in + 1, GMP_NUMB_HIGHBIT);
223   if (UNLIKELY (cy != 0))
224     MPN_ZERO (tp + 1, in);
225   MPN_COPY (ip, tp + 1, in);
226 #endif
227 
228   qh = mpn_preinv_mu_div_qr (qp, rp, np, nn, dp, dn, ip, in, scratch + in);
229 
230   return qh;
231 }
232 
233 mp_limb_t
mpn_preinv_mu_div_qr(mp_ptr qp,mp_ptr rp,mp_srcptr np,mp_size_t nn,mp_srcptr dp,mp_size_t dn,mp_srcptr ip,mp_size_t in,mp_ptr scratch)234 mpn_preinv_mu_div_qr (mp_ptr qp,
235 		      mp_ptr rp,
236 		      mp_srcptr np,
237 		      mp_size_t nn,
238 		      mp_srcptr dp,
239 		      mp_size_t dn,
240 		      mp_srcptr ip,
241 		      mp_size_t in,
242 		      mp_ptr scratch)
243 {
244   mp_size_t qn;
245   mp_limb_t cy, cx, qh;
246   mp_limb_t r;
247   mp_size_t tn, wn;
248 
249 #define tp           scratch
250 #define scratch_out  (scratch + tn)
251 
252   qn = nn - dn;
253 
254   np += qn;
255   qp += qn;
256 
257   qh = mpn_cmp (np, dp, dn) >= 0;
258   if (qh != 0)
259     mpn_sub_n (rp, np, dp, dn);
260   else
261     MPN_COPY_INCR (rp, np, dn);
262 
263   /* if (qn == 0) */			/* The while below handles this case */
264   /*   return qh; */			/* Degenerate use.  Should we allow this? */
265 
266   while (qn > 0)
267     {
268       if (qn < in)
269 	{
270 	  ip += in - qn;
271 	  in = qn;
272 	}
273       np -= in;
274       qp -= in;
275 
276       /* Compute the next block of quotient limbs by multiplying the inverse I
277 	 by the upper part of the partial remainder R.  */
278       mpn_mul_n (tp, rp + dn - in, ip, in);		/* mulhi  */
279       cy = mpn_add_n (qp, tp + in, rp + dn - in, in);	/* I's msb implicit */
280       ASSERT_ALWAYS (cy == 0);
281 
282       qn -= in;
283 
284       /* Compute the product of the quotient block and the divisor D, to be
285 	 subtracted from the partial remainder combined with new limbs from the
286 	 dividend N.  We only really need the low dn+1 limbs.  */
287 
288       if (BELOW_THRESHOLD (in, MUL_TO_MULMOD_BNM1_FOR_2NXN_THRESHOLD))
289 	mpn_mul (tp, dp, dn, qp, in);		/* dn+in limbs, high 'in' cancels */
290       else
291 	{
292 	  tn = mpn_mulmod_bnm1_next_size (dn + 1);
293 	  mpn_mulmod_bnm1 (tp, tn, dp, dn, qp, in, scratch_out);
294 	  wn = dn + in - tn;			/* number of wrapped limbs */
295 	  if (wn > 0)
296 	    {
297 	      cy = mpn_sub_n (tp, tp, rp + dn - wn, wn);
298 	      cy = mpn_sub_1 (tp + wn, tp + wn, tn - wn, cy);
299 	      cx = mpn_cmp (rp + dn - in, tp + dn, tn - dn) < 0;
300 	      ASSERT_ALWAYS (cx >= cy);
301 	      mpn_incr_u (tp, cx - cy);
302 	    }
303 	}
304 
305       r = rp[dn - in] - tp[dn];
306 
307       /* Subtract the product from the partial remainder combined with new
308 	 limbs from the dividend N, generating a new partial remainder R.  */
309       if (dn != in)
310 	{
311 	  cy = mpn_sub_n (tp, np, tp, in);	/* get next 'in' limbs from N */
312 	  cy = mpn_sub_nc (tp + in, rp, tp + in, dn - in, cy);
313 	  MPN_COPY (rp, tp, dn);		/* FIXME: try to avoid this */
314 	}
315       else
316 	{
317 	  cy = mpn_sub_n (rp, np, tp, in);	/* get next 'in' limbs from N */
318 	}
319 
320       STAT (int i; int err = 0;
321 	    static int errarr[5]; static int err_rec; static int tot);
322 
323       /* Check the remainder R and adjust the quotient as needed.  */
324       r -= cy;
325       while (r != 0)
326 	{
327 	  /* We loop 0 times with about 69% probability, 1 time with about 31%
328 	     probability, 2 times with about 0.6% probability, if inverse is
329 	     computed as recommended.  */
330 	  mpn_incr_u (qp, 1);
331 	  cy = mpn_sub_n (rp, rp, dp, dn);
332 	  r -= cy;
333 	  STAT (err++);
334 	}
335       if (mpn_cmp (rp, dp, dn) >= 0)
336 	{
337 	  /* This is executed with about 76% probability.  */
338 	  mpn_incr_u (qp, 1);
339 	  cy = mpn_sub_n (rp, rp, dp, dn);
340 	  STAT (err++);
341 	}
342 
343       STAT (
344 	    tot++;
345 	    errarr[err]++;
346 	    if (err > err_rec)
347 	      err_rec = err;
348 	    if (tot % 0x10000 == 0)
349 	      {
350 		for (i = 0; i <= err_rec; i++)
351 		  printf ("  %d(%.1f%%)", errarr[i], 100.0*errarr[i]/tot);
352 		printf ("\n");
353 	      }
354 	    );
355     }
356 
357   return qh;
358 }
359 
360 /* In case k=0 (automatic choice), we distinguish 3 cases:
361    (a) dn < qn:         in = ceil(qn / ceil(qn/dn))
362    (b) dn/3 < qn <= dn: in = ceil(qn / 2)
363    (c) qn < dn/3:       in = qn
364    In all cases we have in <= dn.
365  */
366 static mp_size_t
mpn_mu_div_qr_choose_in(mp_size_t qn,mp_size_t dn,int k)367 mpn_mu_div_qr_choose_in (mp_size_t qn, mp_size_t dn, int k)
368 {
369   mp_size_t in;
370 
371   if (k == 0)
372     {
373       mp_size_t b;
374       if (qn > dn)
375 	{
376 	  /* Compute an inverse size that is a nice partition of the quotient.  */
377 	  b = (qn - 1) / dn + 1;	/* ceil(qn/dn), number of blocks */
378 	  in = (qn - 1) / b + 1;	/* ceil(qn/b) = ceil(qn / ceil(qn/dn)) */
379 	}
380       else if (3 * qn > dn)
381 	{
382 	  in = (qn - 1) / 2 + 1;	/* b = 2 */
383 	}
384       else
385 	{
386 	  in = (qn - 1) / 1 + 1;	/* b = 1 */
387 	}
388     }
389   else
390     {
391       mp_size_t xn;
392       xn = MIN (dn, qn);
393       in = (xn - 1) / k + 1;
394     }
395 
396   return in;
397 }
398 
399 mp_size_t
mpn_mu_div_qr_itch(mp_size_t nn,mp_size_t dn,int mua_k)400 mpn_mu_div_qr_itch (mp_size_t nn, mp_size_t dn, int mua_k)
401 {
402   mp_size_t in = mpn_mu_div_qr_choose_in (nn - dn, dn, mua_k);
403   mp_size_t itch_preinv = mpn_preinv_mu_div_qr_itch (nn, dn, in);
404   mp_size_t itch_invapp = mpn_invertappr_itch (in + 1) + in + 2; /* 3in + 4 */
405 
406   ASSERT (itch_preinv >= itch_invapp);
407   return in + MAX (itch_invapp, itch_preinv);
408 }
409 
410 mp_size_t
mpn_preinv_mu_div_qr_itch(mp_size_t nn,mp_size_t dn,mp_size_t in)411 mpn_preinv_mu_div_qr_itch (mp_size_t nn, mp_size_t dn, mp_size_t in)
412 {
413   mp_size_t itch_local = mpn_mulmod_bnm1_next_size (dn + 1);
414   mp_size_t itch_out = mpn_mulmod_bnm1_itch (itch_local, dn, in);
415 
416   return itch_local + itch_out;
417 }
418