Since programming the FPGA is very time-consuming, we provide the hardware initialization and hardware run processes independently.
For initializing the hardware:
// Initializes FPGA gqe::FpgaInit init_ocl(xclbin_path); // Allocates pinned host buffers init_ocl.createHostBufs(); // Allocates device buffers init_ocl.createDevBufs();
For loading the table columns and building the bloom-filter’s hash-table:
// Please first load data to corresponding table L column buffers // We assume input table columns is stored in tab_l_col0 & tab_l_col1, // and the corresponding number of rows is stored in table_l_nrow const int BUILD_FACTOR = 10; using namespace xf::database; // Builds 0 - 1/BUILD_FACTOR of table L into bf1 for (int i = 0; i < table_l_nrow / BUILD_FACTOR; i++) { tab_o1_col0[i] = tab_l_col0[i]; tab_o1_col1[i] = tab_l_col1[i]; } // Builds 1/BUILD_FACTOR - 2/BUILD_FACTOR of table L into bf2 for (int i = table_l_nrow / BUILD_FACTOR; i < (table_l_nrow / BUILD_FACTOR) * 2; i++) { tab_o2_col0[i - table_l_nrow / BUILD_FACTOR] = tab_l_col0[i]; tab_o2_col1[i - table_l_nrow / BUILD_FACTOR] = tab_l_col1[i]; } // Total number of unique keys at maximum uint64_t total_num_unique_keys = table_l_nrow / BUILD_FACTOR * 2; // Creates L table gqe::Table tab_l("Table L"); tab.addCol("l_orderkey", gqe::TypeEnum::TypeInt64, tab_l_col0, table_l_nrow); tab.addCol("l_extendedprice", gqe::TypeEnum::TypeInt64, tab_l_col1, table_l_nrow); // Creates O1 table for building bloom-filter-1 gqe::Table tab_o1("Table O1"); tab_o1.addCol("l_orderkey", gqe::TypeEnum::TypeInt64, tab_o1_col0, table_l_nrow / BUILD_FACTOR); tab_o1.addCol("l_extendedprice", gqe::TypeEnum::TypeInt64, tab_o1_col1, table_l_nrow/ BUILD_FACTOR); // Builds bloom-filter-1 gqe::BloomFilter bf1(total_num_unique_keys); bf1.build(tbl_o1, "l_orderkey"); // Creates C table for stroing filtered results, at worst, we'll get every input key passes through the bloom-filter, // so the size of the table c should be the same with table L gqe::Table tab_c("Table C"); tab_c.addCol("c1", gqe::TypeEnum::TypeInt64, tab_c_col0, table_l_nrow); tab_c.addCol("c2", gqe::TypeEnum::TypeInt64, tab_c_col0, table_l_nrow);
Some use cases may need to merge several bloom-filters before running the filtering process:
// Creates O2 table for building bloom-filter-2 gqe::Table tab_o2("Table O2"); tab_o2.addCol("l_orderkey", gqe::TypeEnum::TypeInt64, tab_o2_col0, table_l_nrow / BUILD_FACTOR); tab_o2.addCol("l_extendedprice", gqe::TypeEnum::TypeInt64, tab_o2_col1, table_l_nrow/ BUILD_FACTOR); // Builds bloom-filter-2 gqe::BloomFilter bf2(total_num_unique_keys); bf2.build(tab_o2, "l_orderkey"); // Merges bloom-filter-2 into bloom-filter-1 bf1.merge(bf2);
At last, call the run API of gqe::Filter to perform the pipelined bloom-filtering:
// Creates bloom-filter engine gqe::Filter bigfilter(init_ocl); // Creates StrategySet object to pass on the number of sections of the table gqe::StrategySet params; // if set to 0, the section info should be provided in table L // if set to n (n > 0), the table L should be divided into n sections evenly params.sec_l = 1; gqe::ErrCode err_code; // Performs the bloom-filtering err_code = bigfilter.run(tab_l // input talbe "l_orderkey", // selects key from input table bf1, // bloom-filter which provides hash-table "", // dynamic filter condition, empty for all passes through tab_c, // output table "c1=l_extendedprice, c2=l_orderkey", // output mapping params); // Parameter strcut to provide section info