|
| 1 | +#ifndef TENSORFLOW_CORE_FRAMEWORK_EMBEDDING_CACHE_H_ |
| 2 | +#define TENSORFLOW_CORE_FRAMEWORK_EMBEDDING_CACHE_H_ |
| 3 | +#include <iostream> |
| 4 | +#include <map> |
| 5 | +#include <unordered_map> |
| 6 | +#include <set> |
| 7 | +#include <list> |
| 8 | +#include "tensorflow/core/platform/types.h" |
| 9 | +#include "tensorflow/core/platform/mutex.h" |
| 10 | +#include "tensorflow/core/lib/core/status.h" |
| 11 | + |
| 12 | +namespace tensorflow { |
| 13 | +namespace embedding { |
| 14 | + |
| 15 | +template <class K> |
| 16 | +class BatchCache { |
| 17 | +public: |
| 18 | + BatchCache() {} |
| 19 | + virtual size_t get_evic_ids(K* evic_ids, size_t k_size) = 0; |
| 20 | + virtual void add_to_rank(const K* batch_ids, size_t batch_size) = 0; |
| 21 | + virtual size_t size() = 0; |
| 22 | +}; |
| 23 | + |
| 24 | +template <class K> |
| 25 | +class LRUCache : public BatchCache<K> { |
| 26 | +private: |
| 27 | + class LRUNode { |
| 28 | + public: |
| 29 | + K id; |
| 30 | + LRUNode *pre, *next; |
| 31 | + LRUNode(K id) : id(id), pre(nullptr), next(nullptr) {} |
| 32 | + }; |
| 33 | + LRUNode *head, *tail; |
| 34 | + std::map<K, LRUNode *> mp; |
| 35 | + mutex mu_; |
| 36 | + |
| 37 | +public: |
| 38 | + LRUCache() { |
| 39 | + mp.clear(); |
| 40 | + head = new LRUNode(0); |
| 41 | + tail = new LRUNode(0); |
| 42 | + head->next = tail; |
| 43 | + tail->pre = head; |
| 44 | + } |
| 45 | + |
| 46 | + size_t size() { |
| 47 | + mutex_lock l(mu_); |
| 48 | + return mp.size(); |
| 49 | + } |
| 50 | + |
| 51 | + size_t get_evic_ids(K* evic_ids, size_t k_size) { |
| 52 | + mutex_lock l(mu_); |
| 53 | + size_t true_size = 0; |
| 54 | + LRUNode *evic_node = tail->pre; |
| 55 | + LRUNode *rm_node = evic_node; |
| 56 | + for (size_t i = 0; i < k_size && evic_node != head; ++i) { |
| 57 | + evic_ids[i] = evic_node->id; |
| 58 | + rm_node = evic_node; |
| 59 | + evic_node = evic_node->pre; |
| 60 | + mp.erase(rm_node->id); |
| 61 | + delete rm_node; |
| 62 | + true_size++; |
| 63 | + } |
| 64 | + evic_node->next = tail; |
| 65 | + tail->pre = evic_node; |
| 66 | + return true_size; |
| 67 | + } |
| 68 | + |
| 69 | + void add_to_rank(const K* batch_ids, size_t batch_size) { |
| 70 | + mutex_lock l(mu_); |
| 71 | + for (size_t i = 0; i < batch_size; ++i) { |
| 72 | + K id = batch_ids[i]; |
| 73 | + typename std::map<K, LRUNode *>::iterator it = mp.find(id); |
| 74 | + if (it != mp.end()) { |
| 75 | + LRUNode *node = it->second; |
| 76 | + node->pre->next = node->next; |
| 77 | + node->next->pre = node->pre; |
| 78 | + head->next->pre = node; |
| 79 | + node->next = head->next; |
| 80 | + head->next = node; |
| 81 | + node->pre = head; |
| 82 | + } else { |
| 83 | + LRUNode *newNode = new LRUNode(id); |
| 84 | + head->next->pre = newNode; |
| 85 | + newNode->next = head->next; |
| 86 | + head->next = newNode; |
| 87 | + newNode->pre = head; |
| 88 | + mp[id] = newNode; |
| 89 | + } |
| 90 | + } |
| 91 | + } |
| 92 | +}; |
| 93 | + |
| 94 | +template <class K> |
| 95 | +class LFUCache : public BatchCache<K> { |
| 96 | +private: |
| 97 | + class LFUNode { |
| 98 | + public: |
| 99 | + K key; |
| 100 | + size_t freq; |
| 101 | + LFUNode(K key, size_t freq) : key(key), freq(freq) {} |
| 102 | + }; |
| 103 | + size_t min_freq; |
| 104 | + size_t max_freq; |
| 105 | + std::unordered_map<K, typename std::list<LFUNode>::iterator> key_table; |
| 106 | + std::unordered_map<K, typename std::list<LFUNode>> freq_table; |
| 107 | + mutex mu_; |
| 108 | + |
| 109 | +public: |
| 110 | + LFUCache() { |
| 111 | + min_freq = 0; |
| 112 | + max_freq = 0; |
| 113 | + key_table.clear(); |
| 114 | + freq_table.clear(); |
| 115 | + } |
| 116 | + |
| 117 | + size_t size() { |
| 118 | + mutex_lock l(mu_); |
| 119 | + return key_table.size(); |
| 120 | + } |
| 121 | + |
| 122 | + size_t get_evic_ids(K *evic_ids, size_t k_size) { |
| 123 | + mutex_lock l(mu_); |
| 124 | + size_t true_size = 0; |
| 125 | + for (size_t i = 0; i < k_size; ++i) { |
| 126 | + auto rm_it = freq_table[min_freq].back(); |
| 127 | + key_table.erase(rm_it.key); |
| 128 | + evic_ids[i] = rm_it.key; |
| 129 | + ++true_size; |
| 130 | + freq_table[min_freq].pop_back(); |
| 131 | + if (freq_table[min_freq].size() == 0) { |
| 132 | + freq_table.erase(min_freq); |
| 133 | + ++min_freq; |
| 134 | + while (min_freq <= max_freq) { |
| 135 | + auto it = freq_table.find(min_freq); |
| 136 | + if (it == freq_table.end() || it->second.size() == 0) { |
| 137 | + ++min_freq; |
| 138 | + } else { |
| 139 | + break; |
| 140 | + } |
| 141 | + } |
| 142 | + } |
| 143 | + } |
| 144 | + return true_size; |
| 145 | + } |
| 146 | + |
| 147 | + void add_to_rank(const K *batch_ids, size_t batch_size) { |
| 148 | + mutex_lock l(mu_); |
| 149 | + for (size_t i = 0; i < batch_size; ++i) { |
| 150 | + K id = batch_ids[i]; |
| 151 | + auto it = key_table.find(id); |
| 152 | + if (it == key_table.end()) { |
| 153 | + freq_table[1].push_front(LFUNode(id, 1)); |
| 154 | + key_table[id] = freq_table[1].begin(); |
| 155 | + min_freq = 1; |
| 156 | + } else { |
| 157 | + typename std::list<LFUNode>::iterator node = it->second; |
| 158 | + size_t freq = node->freq; |
| 159 | + freq_table[freq].erase(node); |
| 160 | + if (freq_table[freq].size() == 0) { |
| 161 | + freq_table.erase(freq); |
| 162 | + if (min_freq == freq) |
| 163 | + min_freq += 1; |
| 164 | + } |
| 165 | + max_freq = std::max(max_freq, freq + 1); |
| 166 | + freq_table[freq + 1].push_front(LFUNode(id, freq + 1)); |
| 167 | + key_table[id] = freq_table[freq + 1].begin(); |
| 168 | + } |
| 169 | + } |
| 170 | + } |
| 171 | +}; |
| 172 | + |
| 173 | +} // embedding |
| 174 | +} // tensorflow |
| 175 | + |
| 176 | +#endif // TENSORFLOW_CORE_FRAMEWORK_EMBEDDING_CACHE_H_ |
0 commit comments