a possible low-level optimization

http://www.1point3acres.com/bbs/thread-212960-1-1.html

第二轮白人小哥，一开始问了一道至今不懂的问题，好像是给一个vector<uint8_t> nums, 然后又给一个256位的vector<int> counts，遍历nums，然后counts[nums]++，问如何进行优化，提示说要用到CPU cache之类的东西(完全不知道)。小白哥见我懵逼，后来又给了一道3sum，迅速做出。

uint8_t input[102400];
uint32_t count[256];
void count_it()
{
    for (int i = 0; i < sizeof(input) / sizeof(input[0]); i++) {
        ++count[input[i]];
    }
}

how to optimize? possible points to consider:

a) target "count" array size is 4B*256=1KB, which can fit into L1 cache, so no need to worry about that;

b) input array access is sequential, which is actually cache friendly;

c) update to "count" could have false sharing, but given it‘s all in L1 cache, that‘s fine;

d) optimization 1: the loop could be unrolled to reduce loop check;

e) optimization 2: input array could be pre-fetched (i.e. insert PREFETCH instructions beforehand);

    for (int i = 0; i < sizeof(input) / sizeof(input[0]);) {
        // typical cache size is 64 bytes
        __builtin_prefetch(&input[i+64], 0, 3); // prefetch for read, high locality
        for (int j = 0; j < 8; j++) {
            int k = i + j * 8;
            ++count[input[k]];
            ++count[input[k+1]];
            ++count[input[k+2]];
            ++count[input[k+3]];
            ++count[input[k+4]];
            ++count[input[k+5]];
            ++count[input[k+6]];
            ++count[input[k+7]];
        }
        i += 64;
    }

(see https://gcc.gnu.org/onlinedocs/gcc-5.4.0/gcc/Other-Builtins.html for __builtin_prefetch)

f) optimization 3: multi-threading, but need to use lock instruction when incrementing the count;

g) optimization 4: vector extension CPU instructions: "gather" instruction to load sparse locations (count[xxx]) to a zmmx register (512bit, 64byte i.e. 16 integers), then it can process 16 input uchar8_t in one go; then add a constant 512bit integer which adds 1 to each integer. corresponding "scatter" instruction will store back the updated count.

第二轮白人小哥，一开始问了一道至今不懂的问题，好像是给一个vector<uint8_t> nums, 然后又给一个256位的vector<int> counts，遍历nums，然后counts[nums]++，问如何进行优化，提示说要用到CPU cache之类的东西(完全不知道)。小白哥见我懵逼，后来又给了一道3sum，迅速做出。