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. */
32 /* Overly cautious check for overflow. */
33 if (sizeof(*tally) * buckets / sizeof(*tally) != buckets) {
37 tally = (struct tally *)malloc(
38 sizeof(*tally) + sizeof(tally->counts[0])*(buckets-1));
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;
48 memset(tally->counts, 0, sizeof(tally->counts[0])*buckets);
52 static unsigned bucket_of(ssize_t min, unsigned step_bits, ssize_t val)
54 /* Don't over-shift. */
55 if (step_bits == SIZET_BITS)
57 assert(step_bits < SIZET_BITS);
58 return (size_t)(val - min) >> step_bits;
61 /* Return the min value in bucket b. */
62 static ssize_t bucket_min(ssize_t min, unsigned step_bits, unsigned b)
64 /* Don't over-shift. */
65 if (step_bits == SIZET_BITS)
67 assert(step_bits < SIZET_BITS);
68 return min + ((ssize_t)b << step_bits);
71 /* Does shifting by this many bits truncate the number? */
72 static bool shift_overflows(size_t num, unsigned bits)
77 return ((num << bits) >> 1) != (num << (bits - 1));
80 /* When min or max change, we may need to shuffle the frequency counts. */
81 static void renormalize(struct tally *tally,
82 ssize_t new_min, ssize_t new_max)
85 unsigned int i, old_min;
87 /* Uninitialized? Don't do anything... */
88 if (tally->max < tally->min)
91 /* If we don't have sufficient range, increase step bits until
92 * buckets cover entire range of ssize_t anyway. */
93 range = (new_max - new_min) + 1;
94 while (!shift_overflows(tally->buckets, tally->step_bits)
95 && range > ((size_t)tally->buckets << tally->step_bits)) {
97 for (i = 1; i < tally->buckets; i++) {
98 tally->counts[i/2] += tally->counts[i];
104 /* Now if minimum has dropped, move buckets up. */
105 old_min = bucket_of(new_min, tally->step_bits, tally->min);
106 memmove(tally->counts + old_min,
108 sizeof(tally->counts[0]) * (tally->buckets - old_min));
109 memset(tally->counts, 0, sizeof(tally->counts[0]) * old_min);
111 /* If we moved boundaries, adjust buckets to that ratio. */
112 spill = (tally->min - new_min) % (1 << tally->step_bits);
113 for (i = 0; i < tally->buckets-1; i++) {
114 size_t adjust = (tally->counts[i] >> tally->step_bits) * spill;
115 tally->counts[i] -= adjust;
116 tally->counts[i+1] += adjust;
120 tally->min = new_min;
121 tally->max = new_max;
124 void tally_add(struct tally *tally, ssize_t val)
126 ssize_t new_min = tally->min, new_max = tally->max;
127 bool need_renormalize = false;
129 if (val < tally->min) {
131 need_renormalize = true;
133 if (val > tally->max) {
135 need_renormalize = true;
137 if (need_renormalize)
138 renormalize(tally, new_min, new_max);
140 /* 128-bit arithmetic! If we didn't want exact mean, we could just
141 * pull it out of counts. */
142 if (val > 0 && tally->total[0] + val < tally->total[0])
144 else if (val < 0 && tally->total[0] + val > tally->total[0])
146 tally->total[0] += val;
147 tally->counts[bucket_of(tally->min, tally->step_bits, val)]++;
150 size_t tally_num(const struct tally *tally)
153 for (i = 0; i < tally->buckets; i++)
154 num += tally->counts[i];
158 ssize_t tally_min(const struct tally *tally)
163 ssize_t tally_max(const struct tally *tally)
168 /* FIXME: Own ccan module please! */
169 static unsigned fls64(uint64_t val)
171 #if HAVE_BUILTIN_CLZL
172 if (val <= ULONG_MAX) {
173 /* This is significantly faster! */
174 return val ? sizeof(long) * CHAR_BIT - __builtin_clzl(val) : 0;
181 if (!(val & 0xffffffff00000000ull)) {
185 if (!(val & 0xffff000000000000ull)) {
189 if (!(val & 0xff00000000000000ull)) {
193 if (!(val & 0xf000000000000000ull)) {
197 if (!(val & 0xc000000000000000ull)) {
201 if (!(val & 0x8000000000000000ull)) {
206 #if HAVE_BUILTIN_CLZL
211 /* This is stolen straight from Hacker's Delight. */
212 static uint64_t divlu64(uint64_t u1, uint64_t u0, uint64_t v)
214 const uint64_t b = 4294967296ULL; // Number base (32 bits).
215 uint32_t un[4], // Dividend and divisor
216 vn[2]; // normalized and broken
217 // up into halfwords.
218 uint32_t q[2]; // Quotient as halfwords.
219 uint64_t un1, un0, // Dividend and divisor
220 vn0; // as fullwords.
221 uint64_t qhat; // Estimated quotient digit.
222 uint64_t rhat; // A remainder.
223 uint64_t p; // Product of two digits.
224 int64_t s, i, j, t, k;
226 if (u1 >= v) // If overflow, return the largest
227 return (uint64_t)-1; // possible quotient.
229 s = 64 - fls64(v); // 0 <= s <= 63.
230 vn0 = v << s; // Normalize divisor.
231 vn[1] = vn0 >> 32; // Break divisor up into
232 vn[0] = vn0 & 0xFFFFFFFF; // two 32-bit halves.
234 // Shift dividend left.
235 un1 = ((u1 << s) | (u0 >> (64 - s))) & (-s >> 63);
237 un[3] = un1 >> 32; // Break dividend up into
238 un[2] = un1; // four 32-bit halfwords
239 un[1] = un0 >> 32; // Note: storing into
240 un[0] = un0; // halfwords truncates.
242 for (j = 1; j >= 0; j--) {
243 // Compute estimate qhat of q[j].
244 qhat = (un[j+2]*b + un[j+1])/vn[1];
245 rhat = (un[j+2]*b + un[j+1]) - qhat*vn[1];
247 if (qhat >= b || qhat*vn[0] > b*rhat + un[j]) {
250 if (rhat < b) goto again;
253 // Multiply and subtract.
255 for (i = 0; i < 2; i++) {
257 t = un[i+j] - k - (p & 0xFFFFFFFF);
259 k = (p >> 32) - (t >> 32);
264 q[j] = qhat; // Store quotient digit.
265 if (t < 0) { // If we subtracted too
266 q[j] = q[j] - 1; // much, add back.
268 for (i = 0; i < 2; i++) {
269 t = un[i+j] + vn[i] + k;
273 un[j+2] = un[j+2] + k;
277 return q[1]*b + q[0];
280 static int64_t divls64(int64_t u1, uint64_t u0, int64_t v)
282 int64_t q, uneg, vneg, diff, borrow;
284 uneg = u1 >> 63; // -1 if u < 0.
285 if (uneg) { // Compute the absolute
286 u0 = -u0; // value of the dividend u.
291 vneg = v >> 63; // -1 if v < 0.
292 v = (v ^ vneg) - vneg; // Absolute value of v.
294 if ((uint64_t)u1 >= (uint64_t)v)
297 q = divlu64(u1, u0, v);
299 diff = uneg ^ vneg; // Negate q if signs of
300 q = (q ^ diff) - diff; // u and v differed.
302 if ((diff ^ q) < 0 && q != 0) { // If overflow, return the largest
303 overflow: // possible neg. quotient.
304 q = 0x8000000000000000ULL;
309 ssize_t tally_mean(const struct tally *tally)
311 size_t count = tally_num(tally);
315 if (sizeof(tally->total[0]) == sizeof(uint32_t)) {
316 /* Use standard 64-bit arithmetic. */
317 int64_t total = tally->total[0]
318 | (((uint64_t)tally->total[1]) << 32);
319 return total / count;
321 return divls64(tally->total[1], tally->total[0], count);
324 ssize_t tally_total(const struct tally *tally, ssize_t *overflow)
327 *overflow = tally->total[1];
328 return tally->total[0];
331 /* If result is negative, make sure we can represent it. */
332 if (tally->total[1] & ((size_t)1 << (SIZET_BITS-1))) {
333 /* Must have only underflowed once, and must be able to
334 * represent result at ssize_t. */
335 if ((~tally->total[1])+1 != 0
336 || (ssize_t)tally->total[0] >= 0) {
337 /* Underflow, return minimum. */
338 return (ssize_t)((size_t)1 << (SIZET_BITS - 1));
341 /* Result is positive, must not have overflowed, and must be
342 * able to represent as ssize_t. */
343 if (tally->total[1] || (ssize_t)tally->total[0] < 0) {
344 /* Overflow. Return maximum. */
345 return (ssize_t)~((size_t)1 << (SIZET_BITS - 1));
348 return tally->total[0];
351 static ssize_t bucket_range(const struct tally *tally, unsigned b, size_t *err)
355 min = bucket_min(tally->min, tally->step_bits, b);
356 if (b == tally->buckets - 1)
359 max = bucket_min(tally->min, tally->step_bits, b+1) - 1;
361 /* FIXME: Think harder about cumulative error; is this enough?. */
362 *err = (max - min + 1) / 2;
363 /* Avoid overflow. */
364 return min + (max - min) / 2;
367 ssize_t tally_approx_median(const struct tally *tally, size_t *err)
369 size_t count = tally_num(tally), total = 0;
372 for (i = 0; i < tally->buckets; i++) {
373 total += tally->counts[i];
374 if (total * 2 >= count)
377 return bucket_range(tally, i, err);
380 ssize_t tally_approx_mode(const struct tally *tally, size_t *err)
382 unsigned int i, min_best = 0, max_best = 0;
384 for (i = 0; i < tally->buckets; i++) {
385 if (tally->counts[i] > tally->counts[min_best]) {
386 min_best = max_best = i;
387 } else if (tally->counts[i] == tally->counts[min_best]) {
392 /* We can have more than one best, making our error huge. */
393 if (min_best != max_best) {
395 min = bucket_range(tally, min_best, err);
396 max = bucket_range(tally, max_best, err);
398 *err += (size_t)(max - min);
399 return min + (max - min) / 2;
402 return bucket_range(tally, min_best, err);
405 static unsigned get_max_bucket(const struct tally *tally)
409 for (i = tally->buckets; i > 0; i--)
410 if (tally->counts[i-1])
415 char *tally_histogram(const struct tally *tally,
416 unsigned width, unsigned height)
418 unsigned int i, count, max_bucket, largest_bucket;
422 assert(width >= TALLY_MIN_HISTO_WIDTH);
423 assert(height >= TALLY_MIN_HISTO_HEIGHT);
425 /* Ignore unused buckets. */
426 max_bucket = get_max_bucket(tally);
428 /* FIXME: It'd be nice to smooth here... */
429 if (height >= max_bucket) {
433 /* We create a temporary then renormalize so < height. */
434 /* FIXME: Antialias properly! */
435 tmp = tally_new(tally->buckets);
438 tmp->min = tally->min;
439 tmp->max = tally->max;
440 tmp->step_bits = tally->step_bits;
441 memcpy(tmp->counts, tally->counts,
442 sizeof(tally->counts[0]) * tmp->buckets);
443 while ((max_bucket = get_max_bucket(tmp)) >= height)
444 renormalize(tmp, tmp->min, tmp->max * 2);
446 tmp->max = tally->max;
451 /* Figure out longest line, for scale. */
453 for (i = 0; i < tally->buckets; i++) {
454 if (tally->counts[i] > largest_bucket)
455 largest_bucket = tally->counts[i];
458 p = graph = (char *)malloc(height * (width + 1) + 1);
464 for (i = 0; i < height; i++) {
465 unsigned covered = 1, row;
467 /* People expect minimum at the bottom. */
468 row = height - i - 1;
469 count = (double)tally->counts[row] / largest_bucket * (width-1)+1;
472 covered = snprintf(p, width, "%zi", tally->min);
473 else if (row == height - 1)
474 covered = snprintf(p, width, "%zi", tally->max);
475 else if (row == bucket_of(tally->min, tally->step_bits, 0))
489 memset(p, '*', count);