tally: Adapt renormalize to Samba coding conventions
[ccan] / ccan / tally / tally.c
1 /* Licensed under LGPLv3+ - see LICENSE file for details */
2 #include <ccan/tally/tally.h>
3 #include <ccan/build_assert/build_assert.h>
4 #include <ccan/likely/likely.h>
5 #include <stdint.h>
6 #include <limits.h>
7 #include <string.h>
8 #include <stdio.h>
9 #include <assert.h>
10 #include <stdlib.h>
11
12 #define SIZET_BITS (sizeof(size_t)*CHAR_BIT)
13
14 /* We use power of 2 steps.  I tried being tricky, but it got buggy. */
15 struct tally {
16         ssize_t min, max;
17         size_t total[2];
18         /* This allows limited frequency analysis. */
19         unsigned buckets, step_bits;
20         size_t counts[1 /* Actually: [buckets] */ ];
21 };
22
23 struct tally *tally_new(unsigned buckets)
24 {
25         struct tally *tally;
26
27         /* There is always 1 bucket. */
28         if (buckets == 0) {
29                 buckets = 1;
30         }
31
32         /* Overly cautious check for overflow. */
33         if (sizeof(*tally) * buckets / sizeof(*tally) != buckets) {
34                 return NULL;
35         }
36
37         tally = (struct tally *)malloc(
38                 sizeof(*tally) + sizeof(tally->counts[0])*(buckets-1));
39         if (tally == NULL) {
40                 return NULL;
41         }
42
43         tally->max = ((size_t)1 << (SIZET_BITS - 1));
44         tally->min = ~tally->max;
45         tally->total[0] = tally->total[1] = 0;
46         tally->buckets = buckets;
47         tally->step_bits = 0;
48         memset(tally->counts, 0, sizeof(tally->counts[0])*buckets);
49         return tally;
50 }
51
52 static unsigned bucket_of(ssize_t min, unsigned step_bits, ssize_t val)
53 {
54         /* Don't over-shift. */
55         if (step_bits == SIZET_BITS) {
56                 return 0;
57         }
58         assert(step_bits < SIZET_BITS);
59         return (size_t)(val - min) >> step_bits;
60 }
61
62 /* Return the min value in bucket b. */
63 static ssize_t bucket_min(ssize_t min, unsigned step_bits, unsigned b)
64 {
65         /* Don't over-shift. */
66         if (step_bits == SIZET_BITS) {
67                 return min;
68         }
69         assert(step_bits < SIZET_BITS);
70         return min + ((ssize_t)b << step_bits);
71 }
72
73 /* Does shifting by this many bits truncate the number? */
74 static bool shift_overflows(size_t num, unsigned bits)
75 {
76         if (bits == 0) {
77                 return false;
78         }
79
80         return ((num << bits) >> 1) != (num << (bits - 1));
81 }
82
83 /* When min or max change, we may need to shuffle the frequency counts. */
84 static void renormalize(struct tally *tally,
85                         ssize_t new_min, ssize_t new_max)
86 {
87         size_t range, spill;
88         unsigned int i, old_min;
89
90         /* Uninitialized?  Don't do anything... */
91         if (tally->max < tally->min) {
92                 goto update;
93         }
94
95         /* If we don't have sufficient range, increase step bits until
96          * buckets cover entire range of ssize_t anyway. */
97         range = (new_max - new_min) + 1;
98         while (!shift_overflows(tally->buckets, tally->step_bits)
99                && range > ((size_t)tally->buckets << tally->step_bits)) {
100                 /* Collapse down. */
101                 for (i = 1; i < tally->buckets; i++) {
102                         tally->counts[i/2] += tally->counts[i];
103                         tally->counts[i] = 0;
104                 }
105                 tally->step_bits++;
106         }
107
108         /* Now if minimum has dropped, move buckets up. */
109         old_min = bucket_of(new_min, tally->step_bits, tally->min);
110         memmove(tally->counts + old_min,
111                 tally->counts,
112                 sizeof(tally->counts[0]) * (tally->buckets - old_min));
113         memset(tally->counts, 0, sizeof(tally->counts[0]) * old_min);
114
115         /* If we moved boundaries, adjust buckets to that ratio. */
116         spill = (tally->min - new_min) % (1 << tally->step_bits);
117         for (i = 0; i < tally->buckets-1; i++) {
118                 size_t adjust = (tally->counts[i] >> tally->step_bits) * spill;
119                 tally->counts[i] -= adjust;
120                 tally->counts[i+1] += adjust;
121         }
122
123 update:
124         tally->min = new_min;
125         tally->max = new_max;
126 }
127
128 void tally_add(struct tally *tally, ssize_t val)
129 {
130         ssize_t new_min = tally->min, new_max = tally->max;
131         bool need_renormalize = false;
132
133         if (val < tally->min) {
134                 new_min = val;
135                 need_renormalize = true;
136         }
137         if (val > tally->max) {
138                 new_max = val;
139                 need_renormalize = true;
140         }
141         if (need_renormalize)
142                 renormalize(tally, new_min, new_max);
143
144         /* 128-bit arithmetic!  If we didn't want exact mean, we could just
145          * pull it out of counts. */
146         if (val > 0 && tally->total[0] + val < tally->total[0])
147                 tally->total[1]++;
148         else if (val < 0 && tally->total[0] + val > tally->total[0])
149                 tally->total[1]--;
150         tally->total[0] += val;
151         tally->counts[bucket_of(tally->min, tally->step_bits, val)]++;
152 }
153
154 size_t tally_num(const struct tally *tally)
155 {
156         size_t i, num = 0;
157         for (i = 0; i < tally->buckets; i++)
158                 num += tally->counts[i];
159         return num;
160 }
161
162 ssize_t tally_min(const struct tally *tally)
163 {
164         return tally->min;
165 }
166
167 ssize_t tally_max(const struct tally *tally)
168 {
169         return tally->max;
170 }
171
172 /* FIXME: Own ccan module please! */
173 static unsigned fls64(uint64_t val)
174 {
175 #if HAVE_BUILTIN_CLZL
176         if (val <= ULONG_MAX) {
177                 /* This is significantly faster! */
178                 return val ? sizeof(long) * CHAR_BIT - __builtin_clzl(val) : 0;
179         } else {
180 #endif
181         uint64_t r = 64;
182
183         if (!val)
184                 return 0;
185         if (!(val & 0xffffffff00000000ull)) {
186                 val <<= 32;
187                 r -= 32;
188         }
189         if (!(val & 0xffff000000000000ull)) {
190                 val <<= 16;
191                 r -= 16;
192         }
193         if (!(val & 0xff00000000000000ull)) {
194                 val <<= 8;
195                 r -= 8;
196         }
197         if (!(val & 0xf000000000000000ull)) {
198                 val <<= 4;
199                 r -= 4;
200         }
201         if (!(val & 0xc000000000000000ull)) {
202                 val <<= 2;
203                 r -= 2;
204         }
205         if (!(val & 0x8000000000000000ull)) {
206                 val <<= 1;
207                 r -= 1;
208         }
209         return r;
210 #if HAVE_BUILTIN_CLZL
211         }
212 #endif
213 }
214
215 /* This is stolen straight from Hacker's Delight. */
216 static uint64_t divlu64(uint64_t u1, uint64_t u0, uint64_t v)
217 {
218         const uint64_t b = 4294967296ULL; // Number base (32 bits).
219         uint32_t un[4],           // Dividend and divisor
220                 vn[2];            // normalized and broken
221                                   // up into halfwords.
222         uint32_t q[2];            // Quotient as halfwords.
223         uint64_t un1, un0,        // Dividend and divisor
224                 vn0;              // as fullwords.
225         uint64_t qhat;            // Estimated quotient digit.
226         uint64_t rhat;            // A remainder.
227         uint64_t p;               // Product of two digits.
228         int64_t s, i, j, t, k;
229
230         if (u1 >= v)              // If overflow, return the largest
231                 return (uint64_t)-1; // possible quotient.
232
233         s = 64 - fls64(v);                // 0 <= s <= 63.
234         vn0 = v << s;             // Normalize divisor.
235         vn[1] = vn0 >> 32;        // Break divisor up into
236         vn[0] = vn0 & 0xFFFFFFFF; // two 32-bit halves.
237
238         // Shift dividend left.
239         un1 = ((u1 << s) | (u0 >> (64 - s))) & (-s >> 63);
240         un0 = u0 << s;
241         un[3] = un1 >> 32;        // Break dividend up into
242         un[2] = un1;              // four 32-bit halfwords
243         un[1] = un0 >> 32;        // Note: storing into
244         un[0] = un0;              // halfwords truncates.
245
246         for (j = 1; j >= 0; j--) {
247                 // Compute estimate qhat of q[j].
248                 qhat = (un[j+2]*b + un[j+1])/vn[1];
249                 rhat = (un[j+2]*b + un[j+1]) - qhat*vn[1];
250         again:
251                 if (qhat >= b || qhat*vn[0] > b*rhat + un[j]) {
252                         qhat = qhat - 1;
253                         rhat = rhat + vn[1];
254                         if (rhat < b) goto again;
255                 }
256
257                 // Multiply and subtract.
258                 k = 0;
259                 for (i = 0; i < 2; i++) {
260                         p = qhat*vn[i];
261                         t = un[i+j] - k - (p & 0xFFFFFFFF);
262                         un[i+j] = t;
263                         k = (p >> 32) - (t >> 32);
264                 }
265                 t = un[j+2] - k;
266                 un[j+2] = t;
267
268                 q[j] = qhat;              // Store quotient digit.
269                 if (t < 0) {              // If we subtracted too
270                         q[j] = q[j] - 1;  // much, add back.
271                         k = 0;
272                         for (i = 0; i < 2; i++) {
273                                 t = un[i+j] + vn[i] + k;
274                                 un[i+j] = t;
275                                 k = t >> 32;
276                         }
277                         un[j+2] = un[j+2] + k;
278                 }
279         } // End j.
280
281         return q[1]*b + q[0];
282 }
283
284 static int64_t divls64(int64_t u1, uint64_t u0, int64_t v)
285 {
286         int64_t q, uneg, vneg, diff, borrow;
287
288         uneg = u1 >> 63;          // -1 if u < 0.
289         if (uneg) {               // Compute the absolute
290                 u0 = -u0;         // value of the dividend u.
291                 borrow = (u0 != 0);
292                 u1 = -u1 - borrow;
293         }
294
295         vneg = v >> 63;           // -1 if v < 0.
296         v = (v ^ vneg) - vneg;    // Absolute value of v.
297
298         if ((uint64_t)u1 >= (uint64_t)v)
299                 goto overflow;
300
301         q = divlu64(u1, u0, v);
302
303         diff = uneg ^ vneg;       // Negate q if signs of
304         q = (q ^ diff) - diff;    // u and v differed.
305
306         if ((diff ^ q) < 0 && q != 0) {    // If overflow, return the largest
307         overflow:                          // possible neg. quotient.
308                 q = 0x8000000000000000ULL;
309         }
310         return q;
311 }
312
313 ssize_t tally_mean(const struct tally *tally)
314 {
315         size_t count = tally_num(tally);
316         if (!count)
317                 return 0;
318
319         if (sizeof(tally->total[0]) == sizeof(uint32_t)) {
320                 /* Use standard 64-bit arithmetic. */
321                 int64_t total = tally->total[0]
322                         | (((uint64_t)tally->total[1]) << 32);
323                 return total / count;
324         }
325         return divls64(tally->total[1], tally->total[0], count);
326 }
327
328 ssize_t tally_total(const struct tally *tally, ssize_t *overflow)
329 {
330         if (overflow) {
331                 *overflow = tally->total[1];
332                 return tally->total[0];
333         }
334
335         /* If result is negative, make sure we can represent it. */
336         if (tally->total[1] & ((size_t)1 << (SIZET_BITS-1))) {
337                 /* Must have only underflowed once, and must be able to
338                  * represent result at ssize_t. */
339                 if ((~tally->total[1])+1 != 0
340                     || (ssize_t)tally->total[0] >= 0) {
341                         /* Underflow, return minimum. */
342                         return (ssize_t)((size_t)1 << (SIZET_BITS - 1));
343                 }
344         } else {
345                 /* Result is positive, must not have overflowed, and must be
346                  * able to represent as ssize_t. */
347                 if (tally->total[1] || (ssize_t)tally->total[0] < 0) {
348                         /* Overflow.  Return maximum. */
349                         return (ssize_t)~((size_t)1 << (SIZET_BITS - 1));
350                 }
351         }
352         return tally->total[0];
353 }
354
355 static ssize_t bucket_range(const struct tally *tally, unsigned b, size_t *err)
356 {
357         ssize_t min, max;
358
359         min = bucket_min(tally->min, tally->step_bits, b);
360         if (b == tally->buckets - 1)
361                 max = tally->max;
362         else
363                 max = bucket_min(tally->min, tally->step_bits, b+1) - 1;
364
365         /* FIXME: Think harder about cumulative error; is this enough?. */
366         *err = (max - min + 1) / 2;
367         /* Avoid overflow. */
368         return min + (max - min) / 2;
369 }
370
371 ssize_t tally_approx_median(const struct tally *tally, size_t *err)
372 {
373         size_t count = tally_num(tally), total = 0;
374         unsigned int i;
375
376         for (i = 0; i < tally->buckets; i++) {
377                 total += tally->counts[i];
378                 if (total * 2 >= count)
379                         break;
380         }
381         return bucket_range(tally, i, err);
382 }
383
384 ssize_t tally_approx_mode(const struct tally *tally, size_t *err)
385 {
386         unsigned int i, min_best = 0, max_best = 0;
387
388         for (i = 0; i < tally->buckets; i++) {
389                 if (tally->counts[i] > tally->counts[min_best]) {
390                         min_best = max_best = i;
391                 } else if (tally->counts[i] == tally->counts[min_best]) {
392                         max_best = i;
393                 }
394         }
395
396         /* We can have more than one best, making our error huge. */
397         if (min_best != max_best) {
398                 ssize_t min, max;
399                 min = bucket_range(tally, min_best, err);
400                 max = bucket_range(tally, max_best, err);
401                 max += *err;
402                 *err += (size_t)(max - min);
403                 return min + (max - min) / 2;
404         }
405
406         return bucket_range(tally, min_best, err);
407 }
408
409 static unsigned get_max_bucket(const struct tally *tally)
410 {
411         unsigned int i;
412
413         for (i = tally->buckets; i > 0; i--)
414                 if (tally->counts[i-1])
415                         break;
416         return i;
417 }
418
419 char *tally_histogram(const struct tally *tally,
420                       unsigned width, unsigned height)
421 {
422         unsigned int i, count, max_bucket, largest_bucket;
423         struct tally *tmp;
424         char *graph, *p;
425
426         assert(width >= TALLY_MIN_HISTO_WIDTH);
427         assert(height >= TALLY_MIN_HISTO_HEIGHT);
428
429         /* Ignore unused buckets. */
430         max_bucket = get_max_bucket(tally);
431
432         /* FIXME: It'd be nice to smooth here... */
433         if (height >= max_bucket) {
434                 height = max_bucket;
435                 tmp = NULL;
436         } else {
437                 /* We create a temporary then renormalize so < height. */
438                 /* FIXME: Antialias properly! */
439                 tmp = tally_new(tally->buckets);
440                 if (!tmp)
441                         return NULL;
442                 tmp->min = tally->min;
443                 tmp->max = tally->max;
444                 tmp->step_bits = tally->step_bits;
445                 memcpy(tmp->counts, tally->counts,
446                        sizeof(tally->counts[0]) * tmp->buckets);
447                 while ((max_bucket = get_max_bucket(tmp)) >= height)
448                         renormalize(tmp, tmp->min, tmp->max * 2);
449                 /* Restore max */
450                 tmp->max = tally->max;
451                 tally = tmp;
452                 height = max_bucket;
453         }
454
455         /* Figure out longest line, for scale. */
456         largest_bucket = 0;
457         for (i = 0; i < tally->buckets; i++) {
458                 if (tally->counts[i] > largest_bucket)
459                         largest_bucket = tally->counts[i];
460         }
461
462         p = graph = (char *)malloc(height * (width + 1) + 1);
463         if (!graph) {
464                 free(tmp);
465                 return NULL;
466         }
467
468         for (i = 0; i < height; i++) {
469                 unsigned covered = 1, row;
470
471                 /* People expect minimum at the bottom. */
472                 row = height - i - 1;
473                 count = (double)tally->counts[row] / largest_bucket * (width-1)+1;
474
475                 if (row == 0)
476                         covered = snprintf(p, width, "%zi", tally->min);
477                 else if (row == height - 1)
478                         covered = snprintf(p, width, "%zi", tally->max);
479                 else if (row == bucket_of(tally->min, tally->step_bits, 0))
480                         *p = '+';
481                 else
482                         *p = '|';
483
484                 if (covered > width)
485                         covered = width;
486                 p += covered;
487
488                 if (count > covered)
489                         count -= covered;
490                 else
491                         count = 0;
492
493                 memset(p, '*', count);
494                 p += count;
495                 *p = '\n';
496                 p++;
497         }
498         *p = '\0';
499         free(tmp);
500         return graph;
501 }