Vitis Database Library

Vitis Database Library is an open-sourced Vitis library written in C++ for accelerating database applications in a variety of use cases. It now covers two levels of acceleration: the module level and the pre-defined kernel level, and will evolve to offer the third level as pure software APIs working with pre-defined hardware overlays.

• At module level, it provides an optimized hardware implementation of most common relational database execution plan steps, like hash-join and aggregation.
• In kernel level, the post-bitstream-programmable kernel can be used to map a sequence of execution plan steps, without having to compile FPGA binaries for each query.
• The upcoming software API level will wrap the details of offloading acceleration with prebuilt binary (overlay) and allow users to accelerate supported database tasks on Alveo cards without hardware development.

Since all the kernel code is developed in HLS C++ with the permissive Apache 2.0 license, advanced users can easily tailor, optimize or combine with property logic at any levels. Demo/examples of different database acceleration approach are also provided with the library for easy on-boarding.

Library Overview

• Requirements
• Design Flows
• License
• Trademark Notice
• Release Note

User Guide

• Typical Use Cases
• L1 Module User Guide
  • Primitive Overview
  • Primitive APIs in xf::database
  • Primitive Design Internals
  • Query-Specific Acceleration Demo
• L2 GQE Kernel User Guide
  • GQE Overview
  • GQE Kernel Design
  • GQE Kernel APIs
  • GQE Kernel Acceleration Demo
• L3 GQE Overlay User Guide

Benchmark Result

• TPC-H Queries with GQE

Library API Summary

L1 APIs

Library API	Description
aggregate	A group of overloaded aggregate functions, supports SUM, MAX, MIN, MEAN, VARIANCE, COUNT, COUNTNONZERO operation.
bitonicSort	Bitonic sort is a parallel algorithm for sorting.
bfGen	Generate the bloom-filter in on-chip RAM blocks.
bfGenStream	Generate the bloom-filter in on-chip RAM blocks, and emit the vectors through FIFO upon finish.
bfCheck	Check existence of a value using bloom-filter vectors.
combineCol	A group of overloaded functions for combining two to five columns into one.
splitCol	A group of overloaded functions for splitting previously combined column into two to five separate ones.
directGroupAggregate	Group-by aggregation with limited key width.
duplicateCol	Duplicate one column into two columns.
dynamicEval	Dynamic expression evaluation.
dynamicFilter	A group overloaded functions for filtering payloads according to one to four conditions columns pro gamed at during run-time.
groupAggregate	A series of overloaded functions for group-aggregation. Input rows are required to be sorted by grouping key(s).
hashAntiJoin	Hash-Anti-Join primitive.
hashGroupAggregate	Generic hash group aggregate primitive.
hashJoinMPU	Hash-Join primitive, using multiple DDR/HBM buffers.
hashJoinV3	Hash-Join v3 primitive, it is designed for HBM device and performs better in large size of table.
hashBuildProbeV3	Hash-Build-Probe v3 primitive, it can perform hash build and hash probe separately.
hashJoinV4	Hash-Join v4 primitive, using bloom-filter to enhance performance of hash join, designed for HBM device.
hashBuildProbeV4	Hash-Build-Probe v4 primitive, build and probe are separately performed. This primitive is designed for HBM device only.
hashLookup3	Lookup3 algorithm generates 64-bit or 32-bit hash.
hashMultiJoin	Hash-Multi-Join primitive is based on hashJoinV3, and can be programmed at run-time to perform inner, anti or left join.
hashMurmur3	Murmur3 hash algorithm.
hashPartition	Hash-Partition primitive splits a table into partitions of rows based on hash of a selected key.
hashSemiJoin	Hash-Semi-Join primitive is based on hashJoinMPU, but performs semi-join.
insertSort	Insert sort algorithm on chip.
mergeJoin	Merge join algorithm for sorted tables without duplicated keys in the left table.
mergeLeftJoin	Merge left join function for sorted tables, the left table should not have duplicated keys.
mergeSort	Merge sort algorithm.
nestedLoopJoin	Nested loop join.
scanCmpStrCol	Scan multiple string columns in global memory, and compare each of them with a constant string
scanCol	A group of overloaded functions for Scanning 1 to 6 columns as a table from DDR/HBM buffers.
staticEval	A group of overloaded functions for evaluating a compile-time selected expression on each row with one to four columns.

L2 APIs

gqeAggr	GQE aggregate kernel
gqeJoin	GQE join kernel
gqePart	GQE partition kernel