Welcome to CUDA CODE GENeral

A code-generating GPU database system that executes all 22 TPC-H benchmark queries using CUDA. Queries are compiled at runtime using either NVRTC (fast) or NVCC, with kernel code generated through a composable operator framework. CODEGENeral is extendable and allows benchmarking on custom defined schemas and queries.

       .^.
      /_+_\
     ( o_o )        
  __/  (_)  \__    "Elevate your GPU Database!"
 / /|  ___  |\ \    - the CODEGENERAL
| | | {JIT} | | |   
 \ \|  ---  |/ /
  '--'_____'--'

Features

• All 22 TPC-H queries implemented
• Runtime JIT compilation (NVRTC or NVCC)
• Operator-based code generation framework
• Bitmap-based join acceleration
• Shared memory reductions for high-performance aggregations

Requirements

• CUDA Toolkit (tested with CUDA 11+)
• CMake 3.18+
• C++17 compiler
• TPC-H data files (.tbl format)

Building

mkdir build && cd build
cmake ..
make -j8

Running TPC-H Queries

Basic Usage

# Run with TPC-H data directory
./src/tpch_demo /path/to/tpch-dbgen

# Run with random test data (no data directory)
./src/tpch_demo

Options

Option	Description
--nvcc	Use NVCC compiler instead of NVRTC (slower compilation, same kernel performance)
--dp	Use Dynamic Parallelism when generating the code for launching kernels
--coop	Use cooperative groups and fuse generated code into a single kernel
--ms	Compile kernels sequentially
-v, --verbose	Print generated CUDA kernel code
-q <number>	Execute TPC-H only
-o, --output FILE	Write timing results to CSV file
-h, --help	Show help message

Examples

# Run all queries with NVRTC and export timing to CSV
./src/tpch_demo /path/to/tpch-dbgen -o results.csv

# Run with NVCC compiler and verbose output
./src/tpch_demo /path/to/tpch-dbgen --nvcc -v

# Compare NVRTC vs NVCC compilation times
./src/tpch_demo /path/to/tpch-dbgen -o nvrtc_timing.csv
./src/tpch_demo /path/to/tpch-dbgen --nvcc -o nvcc_timing.csv

CSV Output Format

query,compiler,lineitem_count,compilation_ms,kernel_ms,total_ms
Q1,NVRTC,6001215,117.38,30.65,148.04
Q2,NVRTC,6001215,535.31,0.24,535.54
...

Code Generation Framework (codegen)

The query engine uses a composable operator framework to generate CUDA kernels. Operators are chained together using a producer-consumer pattern.

Core Infrastructure

Class	Description
Codegen	Code generator that builds CUDA kernel source code. Manages indentation, parameters, and code blocks.
Operator	Abstract base class for all operators. Defines produce(codegen, consume) interface.
UnaryOperator	Base class for operators with a single child operator.

Table Scan Operators

Operator	Description
GPUTableScan	Basic table scan with simple for-loop. Iterates over tuples with idx variable.
GPUTableScanGridStride	Table scan using grid-stride loop pattern for better GPU occupancy.

Selection (Filter) Operators

Operator	Description
Selection	Filter rows using a predicate (for basic scans). Generates if (predicate) { ... }.

Bitmap Operations

Operator	Description
BitmapBuild	Build a bitmap from qualifying rows: bitmap[keyExpr] = 1.
BitmapJoin	Filter rows by checking bitmap: if (bitmap[keyExpr]) { ... }.
MultiBitmapJoin	Join on multiple bitmaps simultaneously (AND logic).
AntiBitmapJoin	Anti-join: passes rows where bitmap is NOT set: if (!bitmap[keyExpr]) { ... }.
TableBitmapBuild	Standalone operator that scans a table and builds a bitmap in one step.
TableBitmapBuildGridStride	Grid-stride version of TableBitmapBuild.

Array Operations

Operator	Description
ArrayLookup	Read value from array: resultVar = array[keyExpr].
ArrayStore	Write value to array: array[keyExpr] = valueExpr.

Aggregation Operators

Operator	Description
AtomicArrayAgg	Atomic aggregation by key: atomicAdd(&array[bucketExpr], valueExpr).
AtomicArrayCount	Atomic count by key: atomicAdd(&array[keyExpr], 1).
AtomicCount	Atomic increment of a single counter.
SharedMemReductionAgg	High-performance reduction using shared memory (one atomicAdd per block).
ArrayMaxReduction	Find maximum value in an array using shared memory reduction.
KeyedAggregation	GROUP BY aggregation with multiple aggregates per bucket.
KeyedDualAggregation	Dual aggregation with bucket key (GROUP BY with two aggregates).

Expression Operators

Operator	Description
ComputeExpr	Compute a derived value: type varName = expression.

Example: Building a Query Kernel

#include "codegen/codegen.hpp"

// Generate kernel for: SELECT SUM(l_extendedprice) FROM lineitem WHERE l_shipdate >= '1994-01-01'
codegen::Codegen cg;
cg.setKernelName("aggregateRevenue");

auto scan = std::make_unique<codegen::GPUTableScanGridStride>("lineitem", "LineItemTuple", "li");
auto filter = std::make_unique<codegen::Selection>(std::move(scan),
    "date_ge(li.l_shipdate, date_start)");
codegen::AtomicArrayAgg agg(std::move(filter), "d_result", "0", "li.l_extendedprice", "double");

agg.produce(&cg, [](){});
std::string kernelCode = cg.print();

Project Structure

CUDACodeGeneral/
├── codegen/
│   └── codegen.hpp      # Codegen framework
│   └── operator.hpp     # Operator framework
├── queries/tpch/
│   └── tpch_q1.hpp ... tpch_q22.hpp  # TPC-H query implementations
├── schema/tables/tpch/
│   └── tpch_schema.hpp     # TPC-H table definitions
├── src/
│   ├── tpch_demo.cu  # Main executable
│   ├── tpch_loader.hpp     # TPC-H data loader
│   ├── launcher.hpp        # Kernel launch utilities
│   └── jit_compiler.hpp    # NVRTC/NVCC compilation
└── build/                  # Build output

Performance Notes

• NVRTC compilation is typically 3-6x faster than NVCC
• Kernel execution times are identical between compilers
• Shared memory reductions significantly outperform naive atomic aggregations
• Bitmap joins provide efficient multi-table query execution

License

See LICENSE file for details.