Напоминание: в случае технических проблем любого характера, вы можете использовать m1.codeforces.com, m2.codeforces.com, m3.codeforces.com. ×

Easy and (Semi)Efficient Dynamic Segment Trees (with Policy Hash Tables)

Правка en11, от Chilli, 2018-12-10 01:58:10

One of my favorite implementations of segment trees has always been "Easy and Efficient Segment Trees, by Al.Cash. I used to dread segtree problems, but after reading that blog post and adapting a super simple implementation I've gotten a lot better with them. However, there are some types of segtree that you can't implement in that fashion, namely dynamic segtrees and persistent segtrees. See here for criticism. With the advent of policy hash tables, however, one can now implement dynamic segtrees in Al.Cash's style with somewhat comparable performance to a custom dynamic segtree.

Standard segtree

This is how a standard segtree looks like. You can set a single element, and query for ranges. It's nice and simple, and I think it's a great implementation.

int N;
int seg[2 * MAXN];

void modify(int p, int val) {
    for (seg[p += N] = val; p > 0; p >>= 1)
        seg[p >> 1] = seg[p] + seg[p ^ 1];
}

int query(int l, int r) {
    int res = 0;
    for (l += N, r += N; l < r; l >>= 1, r >>= 1) {
        if (l & 1)
            res += seg[l++];
        if (r & 1)
            res += seg[--r];
    }
    return res;
}

Dynamic Segtree

However, say your underlying array had 1e9 possible locations, but it only contained 1e5 elements. For example, take a look at this post. Obviously, you can't store all 2e9 elements in your segtree, so what should you do? Here's one solution, replace the array with a hash table. However, as adamant mentions, unordered_map has too much overhead. We'll be benchmarking against the dynamic segtree provided here. I'll also be using a custom hash function. So to be clear, the implementation now looks like:

Code

And benchmarking it with 1e5 random insertions and 1e5 random queries.

pointer: 0.171485
unordered_map: 2.0646

Wow. The unordered_map is nearly 12x slower. That's not really feasible for a lot of contests. What if we replace it with a policy hash table, though?

Code
pointer: 0.202186
policy hash table: 0.384312

Only a 2x decrease in speed. That's already very feasible. However, one might notice that since maps in C++ create elements if you try to access a key that doesn't exist, we're creating a lot of useless elements. Thus, we can simply wrap a check to make sure the element is in the array before we try to access it

EDIT: Updated with dalex's optimization.

gp_hash_table<ll, ll, chash> seg;

ll get(ll x) { return (seg.find(x) == seg.end()) ? 0 : seg[x]; }
void modify(ll p, ll val) {
    for (seg[p += N] = val; p > 0; p >>= 1) {
        seg[p >> 1] = get(p) + get(p ^ 1);
    }
}

ll query(ll l, ll r) {
    ll res = 0;
    for (l += N, r += N; l < r; l >>= 1, r >>= 1) {
        if (l & 1)
            res += get(l++);
        if (r & 1)
            res += get(--r);
    }
    return res;
}

Results (averaged over twenty runs):

2e5 insertions and 2e5 queries

pointer: 0.44085
policy hash table: 0.57878

1e5 insertions and 1e5 queries

pointer: 0.19855
policy hash table: 0.29467

1e4 insertions and 1e4 queries

pointer: 0.014
policy hash table: 0.027

So while we're nearly twice as slow with 1e4 elements and 1e4 queries, we're actually only 30% slower with 2e5 insertions and 2e5 queries.

One more thing. While I'm giving numbers like "30% slower", that's a little bit misleading. If we break down the numbers between insertion/querying, we see this:

2e5 insertions and 2e5 queries Queries:

pointer: 0.41625
policy hash table: 0.15627

Insertions:

pointer: 0.1367
policy hash table: 0.42619

1e4 insertions and 1e4 queries Queries:

pointer : 0.094
policy hash table: 0.007

Insertions:

pointer : 0.0045
policy hash table: 0.0191

So as we see from this more granular breakdown, the Policy Hash Table implementation is actually ~3x faster at querying than the custom implementation, while the custom implementation is roughly ~3x faster at inserting elements.

TL;DR: Using policy hash tables is an extremely easy and fairly efficient method of implementing dynamic segtrees.

Теги segment tree, policy based, dynamic

История

 
 
 
 
Правки
 
 
  Rev. Язык Кто Когда Δ Комментарий
en11 Английский Chilli 2018-12-10 01:58:10 0 (published)
en10 Английский Chilli 2018-12-10 01:57:34 16 Tiny change: ' queries\n~~~~~\np' -> ' queries\n\n~~~~~\np' (saved to drafts)
en9 Английский Chilli 2018-12-10 01:38:15 1038
en8 Английский Chilli 2018-12-09 10:24:32 0 (published)
en7 Английский Chilli 2018-12-09 10:23:47 18 Tiny change: 'og/entry/19080), unorder' -> 'og/entry/18051?#comment-288074), unorder' (saved to drafts)
en6 Английский Chilli 2018-12-09 10:21:18 212 (published)
en5 Английский Chilli 2018-12-09 10:19:30 6
en4 Английский Chilli 2018-07-27 06:48:45 9 Tiny change: 'tyle with comparabl' -> 'tyle with somewhat comparabl'
en3 Английский Chilli 2018-07-27 06:48:10 7 Tiny change: '="Code">\n\n~~~~~\nint N;\n' -> '="Code">\n~~~~~\n\nint N;\n'
en2 Английский Chilli 2018-07-27 06:46:51 91 Tiny change: 'er:AlCash]'s style w' -> 'er:AlCash] 's style w'
en1 Английский Chilli 2018-07-26 04:18:23 4617 Initial revision (saved to drafts)