bb_segsort (segmented sort): Fast Segmented Sort on GPUs

This repository provides a fast segmented sort on NVIDIA GPUs. The library contains many parallel kernels for different types of segments. In particular, the kernels for solving short/medium segments are automatically generated to efficiently utilize registers in GPUs. More details about the kernels and code generation can be found in the original paper.

Original Work

• Original GitHub repository
• Contact Email: [email protected]

Improvements in this fork

• Added key only version
• Asynchronous execution using a single CUDA stream inside bb_segsort_run
• No temporary memory allocation inside bb_segsort_run
• Reduced memory overhead
• Two dimensional kernel grid to avoid index calculations
• Avoiding boundaries check by using one-past-the-end offset
• No dependency on Thrust

Interface differences

• This version expects two offset arrays, one for begin and one for (one-past-the-)end offsets of the segments
• You can use a single array and pass offsets and offsets+1 if the segments are densly packed (end of a segment is begin of next segment). Be sure to include the last one-past-the-end offset.

Usage

To use the segmented sort (bb_segsort), you need to include the bb_segsort.cuh (key-value) or bb_segsort_keys.cuh (key only). Use bb_segsort(...) if you don't care about memory allocation or asynchronous execution, or use bb_segsort_run(...) and provide your own memory allocation and stream.

Note, bb_segsort utilizes an unstable sorting network as the building block; thus, equivalent elements are not guaranteed to keep the original relative order.

Example

main.cu contains an example of how to use (bb_segsort). Adapt the Makefile to fit your system. Especially, you may need to change the ARCH according to your GPU platform. For example, if you are using a P100 GPU, you should update ARCH to 61.

Compile using make:

$ make

After compilation, run the executable:

$ ./main.out

License

Please refer to the included LICENSE file.