Speed up Bloom filter construction (only for the 100M case so far)

Author: thomasmueller

benchmarks/bulk-insert-and-query.cc

@@ -338,7 +338,7 @@ struct FilterAPI<BloomFilter<ItemType, bits_per_item, HashFamily>> {
     table->Add(key);
   }
   static void AddAll(const vector<ItemType> keys, const size_t start, const size_t end, Table* table) {
-      throw std::runtime_error("Unsupported");
+    table->AddAll(keys, start, end);
   }
 
   CONTAIN_ATTRIBUTES
@@ -807,6 +807,28 @@ int main(int argc, char * argv[]) {
   }
 
 
+  if (algorithmId == 37 || algorithmId < 0) {
+      auto cf = FilterBenchmark<
+          BloomFilter<uint64_t, 8, SimpleMixSplit>>(
+          add_count, to_add, to_lookup, seed, true);
+      cout << setw(NAME_WIDTH) << "Bloom8" << cf << endl;
+  }
+
+  if (algorithmId == 38 || algorithmId < 0) {
+      auto cf = FilterBenchmark<
+          BloomFilter<uint64_t, 12, SimpleMixSplit>>(
+          add_count, to_add, to_lookup, seed, true);
+      cout << setw(NAME_WIDTH) << "Bloom12" << cf << endl;
+  }
+
+  if (algorithmId == 39 || algorithmId < 0) {
+      auto cf = FilterBenchmark<
+          BloomFilter<uint64_t, 16, SimpleMixSplit>>(
+          add_count, to_add, to_lookup, seed, true);
+      cout << setw(NAME_WIDTH) << "Bloom16" << cf << endl;
+  }
+
+
 // broken algorithms (don't always find all key)
 /*
   if (algorithmId == 25) {

src/bloom.h

@@ -60,6 +60,9 @@ class BloomFilter {
   // Add an item to the filter.
   Status Add(const ItemType &item);
 
+  // Add multiple items to the filter.
+  Status AddAll(const vector<ItemType> data, const size_t start, const size_t end);
+
   // Report if the item is inserted, with false positive rate.
   Status Contain(const ItemType &item) const;
 
@@ -92,6 +95,51 @@ Status BloomFilter<ItemType, bits_per_item, HashFamily, k>::Add(
     return Ok;
 }
 
+#define BLOCK_SHIFT 18
+#define BLOCK_LEN (1 << BLOCK_SHIFT)
+
+void applyBlock(uint64_t* tmp, int block, int len, uint64_t *data) {
+    for (int i = 0; i < len; i += 2) {
+        int index = tmp[(block << BLOCK_SHIFT) + i];
+        uint64_t bits = tmp[(block << BLOCK_SHIFT) + i + 1];
+        data[index] |= bits;
+    }
+}
+
+template <typename ItemType, size_t bits_per_item,
+    typename HashFamily, int k>
+Status BloomFilter<ItemType, bits_per_item, HashFamily, k>::AddAll(
+    const vector<ItemType> keys, const size_t start, const size_t end) {
+    int blocks = 1 + arrayLength / BLOCK_LEN;
+    uint64_t* tmp = new uint64_t[blocks * BLOCK_LEN];
+    int* tmpLen = new int[blocks]();
+    for(size_t i = start; i < end; i++) {
+        uint64_t key = keys[i];
+        uint64_t hash = hasher(key);
+        uint32_t a = (uint32_t) (hash >> 32);
+        uint32_t bb = (uint32_t) hash;
+        for (int j = 0; j < k; j++) {
+            int index = reduce(a, this->arrayLength);
+            int block = index >> BLOCK_SHIFT;
+            int len = tmpLen[block];
+            tmp[(block << BLOCK_SHIFT) + len] = index;
+            tmp[(block << BLOCK_SHIFT) + len + 1] = getBit(a);
+            tmpLen[block] = len + 2;
+            if (len + 2 == BLOCK_LEN) {
+                applyBlock(tmp, block, len + 2, data);
+                tmpLen[block] = 0;
+            }
+            a += bb;
+        }
+    }
+    for (int block = 0; block < blocks; block++) {
+        applyBlock(tmp, block, tmpLen[block], data);
+    }
+    delete[] tmp;
+    delete[] tmpLen;
+    return Ok;
+}
+
 template <typename ItemType, size_t bits_per_item,
     typename HashFamily, int k>
 Status BloomFilter<ItemType, bits_per_item, HashFamily, k>::Contain(