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