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>
12 #define SIZET_BITS (sizeof(size_t)*CHAR_BIT)
14 /* We use power of 2 steps. I tried being tricky, but it got buggy. */
18 /* This allows limited frequency analysis. */
19 unsigned buckets, step_bits;
20 size_t counts[1 /* Actually: [buckets] */ ];
23 struct tally *tally_new(unsigned buckets)
27 /* There is always 1 bucket. */
31 /* Overly cautious check for overflow. */
32 if (sizeof(*tally) * buckets / sizeof(*tally) != buckets)
34 tally = malloc(sizeof(*tally) + sizeof(tally->counts[0])*(buckets-1));
36 tally->max = ((size_t)1 << (SIZET_BITS - 1));
37 tally->min = ~tally->max;
38 tally->total[0] = tally->total[1] = 0;
39 tally->buckets = buckets;
41 memset(tally->counts, 0, sizeof(tally->counts[0])*buckets);
46 static unsigned bucket_of(ssize_t min, unsigned step_bits, ssize_t val)
48 /* Don't over-shift. */
49 if (step_bits == SIZET_BITS)
51 assert(step_bits < SIZET_BITS);
52 return (size_t)(val - min) >> step_bits;
55 /* Return the min value in bucket b. */
56 static ssize_t bucket_min(ssize_t min, unsigned step_bits, unsigned b)
58 /* Don't over-shift. */
59 if (step_bits == SIZET_BITS)
61 assert(step_bits < SIZET_BITS);
62 return min + ((ssize_t)b << step_bits);
65 /* Does shifting by this many bits truncate the number? */
66 static bool shift_overflows(size_t num, unsigned bits)
71 return ((num << bits) >> 1) != (num << (bits - 1));
74 /* When min or max change, we may need to shuffle the frequency counts. */
75 static void renormalize(struct tally *tally,
76 ssize_t new_min, ssize_t new_max)
79 unsigned int i, old_min;
81 /* Uninitialized? Don't do anything... */
82 if (tally->max < tally->min)
85 /* If we don't have sufficient range, increase step bits until
86 * buckets cover entire range of ssize_t anyway. */
87 range = (new_max - new_min) + 1;
88 while (!shift_overflows(tally->buckets, tally->step_bits)
89 && range > ((size_t)tally->buckets << tally->step_bits)) {
91 for (i = 1; i < tally->buckets; i++) {
92 tally->counts[i/2] += tally->counts[i];
98 /* Now if minimum has dropped, move buckets up. */
99 old_min = bucket_of(new_min, tally->step_bits, tally->min);
100 memmove(tally->counts + old_min,
102 sizeof(tally->counts[0]) * (tally->buckets - old_min));
103 memset(tally->counts, 0, sizeof(tally->counts[0]) * old_min);
105 /* If we moved boundaries, adjust buckets to that ratio. */
106 spill = (tally->min - new_min) % (1 << tally->step_bits);
107 for (i = 0; i < tally->buckets-1; i++) {
108 size_t adjust = (tally->counts[i] >> tally->step_bits) * spill;
109 tally->counts[i] -= adjust;
110 tally->counts[i+1] += adjust;
114 tally->min = new_min;
115 tally->max = new_max;
118 void tally_add(struct tally *tally, ssize_t val)
120 ssize_t new_min = tally->min, new_max = tally->max;
121 bool need_renormalize = false;
123 if (val < tally->min) {
125 need_renormalize = true;
127 if (val > tally->max) {
129 need_renormalize = true;
131 if (need_renormalize)
132 renormalize(tally, new_min, new_max);
134 /* 128-bit arithmetic! If we didn't want exact mean, we could just
135 * pull it out of counts. */
136 if (val > 0 && tally->total[0] + val < tally->total[0])
138 else if (val < 0 && tally->total[0] + val > tally->total[0])
140 tally->total[0] += val;
141 tally->counts[bucket_of(tally->min, tally->step_bits, val)]++;
144 size_t tally_num(const struct tally *tally)
147 for (i = 0; i < tally->buckets; i++)
148 num += tally->counts[i];
152 ssize_t tally_min(const struct tally *tally)
157 ssize_t tally_max(const struct tally *tally)
162 /* FIXME: Own ccan module please! */
163 static unsigned fls64(uint64_t val)
165 #if HAVE_BUILTIN_CLZL
166 if (val <= ULONG_MAX) {
167 /* This is significantly faster! */
168 return val ? sizeof(long) * CHAR_BIT - __builtin_clzl(val) : 0;
175 if (!(val & 0xffffffff00000000ull)) {
179 if (!(val & 0xffff000000000000ull)) {
183 if (!(val & 0xff00000000000000ull)) {
187 if (!(val & 0xf000000000000000ull)) {
191 if (!(val & 0xc000000000000000ull)) {
195 if (!(val & 0x8000000000000000ull)) {
200 #if HAVE_BUILTIN_CLZL
205 /* This is stolen straight from Hacker's Delight. */
206 static uint64_t divlu64(uint64_t u1, uint64_t u0, uint64_t v)
208 const uint64_t b = 4294967296ULL; // Number base (32 bits).
209 uint32_t un[4], // Dividend and divisor
210 vn[2]; // normalized and broken
211 // up into halfwords.
212 uint32_t q[2]; // Quotient as halfwords.
213 uint64_t un1, un0, // Dividend and divisor
214 vn0; // as fullwords.
215 uint64_t qhat; // Estimated quotient digit.
216 uint64_t rhat; // A remainder.
217 uint64_t p; // Product of two digits.
218 int64_t s, i, j, t, k;
220 if (u1 >= v) // If overflow, return the largest
221 return (uint64_t)-1; // possible quotient.
223 s = 64 - fls64(v); // 0 <= s <= 63.
224 vn0 = v << s; // Normalize divisor.
225 vn[1] = vn0 >> 32; // Break divisor up into
226 vn[0] = vn0 & 0xFFFFFFFF; // two 32-bit halves.
228 // Shift dividend left.
229 un1 = ((u1 << s) | (u0 >> (64 - s))) & (-s >> 63);
231 un[3] = un1 >> 32; // Break dividend up into
232 un[2] = un1; // four 32-bit halfwords
233 un[1] = un0 >> 32; // Note: storing into
234 un[0] = un0; // halfwords truncates.
236 for (j = 1; j >= 0; j--) {
237 // Compute estimate qhat of q[j].
238 qhat = (un[j+2]*b + un[j+1])/vn[1];
239 rhat = (un[j+2]*b + un[j+1]) - qhat*vn[1];
241 if (qhat >= b || qhat*vn[0] > b*rhat + un[j]) {
244 if (rhat < b) goto again;
247 // Multiply and subtract.
249 for (i = 0; i < 2; i++) {
251 t = un[i+j] - k - (p & 0xFFFFFFFF);
253 k = (p >> 32) - (t >> 32);
258 q[j] = qhat; // Store quotient digit.
259 if (t < 0) { // If we subtracted too
260 q[j] = q[j] - 1; // much, add back.
262 for (i = 0; i < 2; i++) {
263 t = un[i+j] + vn[i] + k;
267 un[j+2] = un[j+2] + k;
271 return q[1]*b + q[0];
274 static int64_t divls64(int64_t u1, uint64_t u0, int64_t v)
276 int64_t q, uneg, vneg, diff, borrow;
278 uneg = u1 >> 63; // -1 if u < 0.
279 if (uneg) { // Compute the absolute
280 u0 = -u0; // value of the dividend u.
285 vneg = v >> 63; // -1 if v < 0.
286 v = (v ^ vneg) - vneg; // Absolute value of v.
288 if ((uint64_t)u1 >= (uint64_t)v)
291 q = divlu64(u1, u0, v);
293 diff = uneg ^ vneg; // Negate q if signs of
294 q = (q ^ diff) - diff; // u and v differed.
296 if ((diff ^ q) < 0 && q != 0) { // If overflow, return the largest
297 overflow: // possible neg. quotient.
298 q = 0x8000000000000000ULL;
303 ssize_t tally_mean(const struct tally *tally)
305 size_t count = tally_num(tally);
309 if (sizeof(tally->total[0]) == sizeof(uint32_t)) {
310 /* Use standard 64-bit arithmetic. */
311 int64_t total = tally->total[0]
312 | (((uint64_t)tally->total[1]) << 32);
313 return total / count;
315 return divls64(tally->total[1], tally->total[0], count);
318 ssize_t tally_total(const struct tally *tally, ssize_t *overflow)
321 *overflow = tally->total[1];
322 return tally->total[0];
325 /* If result is negative, make sure we can represent it. */
326 if (tally->total[1] & ((size_t)1 << (SIZET_BITS-1))) {
327 /* Must have only underflowed once, and must be able to
328 * represent result at ssize_t. */
329 if ((~tally->total[1])+1 != 0
330 || (ssize_t)tally->total[0] >= 0) {
331 /* Underflow, return minimum. */
332 return (ssize_t)((size_t)1 << (SIZET_BITS - 1));
335 /* Result is positive, must not have overflowed, and must be
336 * able to represent as ssize_t. */
337 if (tally->total[1] || (ssize_t)tally->total[0] < 0) {
338 /* Overflow. Return maximum. */
339 return (ssize_t)~((size_t)1 << (SIZET_BITS - 1));
342 return tally->total[0];
345 static ssize_t bucket_range(const struct tally *tally, unsigned b, size_t *err)
349 min = bucket_min(tally->min, tally->step_bits, b);
350 if (b == tally->buckets - 1)
353 max = bucket_min(tally->min, tally->step_bits, b+1) - 1;
355 /* FIXME: Think harder about cumulative error; is this enough?. */
356 *err = (max - min + 1) / 2;
357 /* Avoid overflow. */
358 return min + (max - min) / 2;
361 ssize_t tally_approx_median(const struct tally *tally, size_t *err)
363 size_t count = tally_num(tally), total = 0;
366 for (i = 0; i < tally->buckets; i++) {
367 total += tally->counts[i];
368 if (total * 2 >= count)
371 return bucket_range(tally, i, err);
374 ssize_t tally_approx_mode(const struct tally *tally, size_t *err)
376 unsigned int i, min_best = 0, max_best = 0;
378 for (i = 0; i < tally->buckets; i++) {
379 if (tally->counts[i] > tally->counts[min_best]) {
380 min_best = max_best = i;
381 } else if (tally->counts[i] == tally->counts[min_best]) {
386 /* We can have more than one best, making our error huge. */
387 if (min_best != max_best) {
389 min = bucket_range(tally, min_best, err);
390 max = bucket_range(tally, max_best, err);
392 *err += (size_t)(max - min);
393 return min + (max - min) / 2;
396 return bucket_range(tally, min_best, err);
399 static unsigned get_max_bucket(const struct tally *tally)
403 for (i = tally->buckets; i > 0; i--)
404 if (tally->counts[i-1])
409 char *tally_histogram(const struct tally *tally,
410 unsigned width, unsigned height)
412 unsigned int i, count, max_bucket, largest_bucket;
416 assert(width >= TALLY_MIN_HISTO_WIDTH);
417 assert(height >= TALLY_MIN_HISTO_HEIGHT);
419 /* Ignore unused buckets. */
420 max_bucket = get_max_bucket(tally);
422 /* FIXME: It'd be nice to smooth here... */
423 if (height >= max_bucket) {
427 /* We create a temporary then renormalize so < height. */
428 /* FIXME: Antialias properly! */
429 tmp = tally_new(tally->buckets);
432 tmp->min = tally->min;
433 tmp->max = tally->max;
434 tmp->step_bits = tally->step_bits;
435 memcpy(tmp->counts, tally->counts,
436 sizeof(tally->counts[0]) * tmp->buckets);
437 while ((max_bucket = get_max_bucket(tmp)) >= height)
438 renormalize(tmp, tmp->min, tmp->max * 2);
440 tmp->max = tally->max;
445 /* Figure out longest line, for scale. */
447 for (i = 0; i < tally->buckets; i++) {
448 if (tally->counts[i] > largest_bucket)
449 largest_bucket = tally->counts[i];
452 p = graph = malloc(height * (width + 1) + 1);
458 for (i = 0; i < height; i++) {
459 unsigned covered = 1, row;
461 /* People expect minimum at the bottom. */
462 row = height - i - 1;
463 count = (double)tally->counts[row] / largest_bucket * (width-1)+1;
466 covered = snprintf(p, width, "%zi", tally->min);
467 else if (row == height - 1)
468 covered = snprintf(p, width, "%zi", tally->max);
469 else if (row == bucket_of(tally->min, tally->step_bits, 0))
483 memset(p, '*', count);