char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

alex-spacemitggml-cpu: Add IME2 Instruction Support for the SpacemiT Backend (#22863)81b0d882a

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
1007 B34 linesraw
1#pragma once
2
3#include <atomic>
4#include <cstdint>
5
6#define SPINE_CACHE_LINE  64
7#define SPINE_CACHE_ALIGN __attribute__((aligned(SPINE_CACHE_LINE)))
8
9struct spine_barrier_t {
10    SPINE_CACHE_ALIGN std::atomic<int64_t> pending_;
11    SPINE_CACHE_ALIGN std::atomic<int64_t> rounds_;
12    SPINE_CACHE_ALIGN int64_t              total_;
13};
14
15inline void spine_barrier_wait(spine_barrier_t * b) {
16    auto cur_round = b->rounds_.load(std::memory_order_acquire);
17    auto cnt       = --b->pending_;
18    if (cnt == 0) {
19        b->pending_.store(b->total_);
20        b->rounds_.store(cur_round + 1);
21    } else {
22        while (cur_round == b->rounds_.load(std::memory_order_relaxed)) {
23            __asm__ volatile("pause " ::: "memory");
24        }
25    }
26}
27
28inline void spine_barrier_init(spine_barrier_t * b, int num_barriers, uint64_t thread_count) {
29    for (int i = 0; i < num_barriers; i++) {
30        b[i].total_ = thread_count;
31        b[i].pending_.store(thread_count);
32        b[i].rounds_.store(0);
33    }
34}