Thrust / Cub Codebase Structure

[jeromeku]

Trying to understand the architecture of Thrust and Cub. Is there documentation on the structure of the codebase, specifically, how the high level abstractions provided by Thrust and Cub are ultimately transformed into primitive cuda memory allocations / kernel launches? Would appreciate any explanations or pointers.

Thanks!

[pauleonix]

For Thrust you can find most of the interesting CUDA-specific stuff in thrust/system/cuda/detail.

[jeromeku]

Thanks -- I'm looking more for an overview of how the code is organized, and an explanation of the core abstractions (e.g., agent) used to implement the various features of thrust and cub.

[gevtushenko]

Hello @jeromeku!

For CUB you can take a look at the Device-Scope section of the developer overview. It should answer your questions on code organization surrounding CUB kernels.

I don't think we have a developer overview for Thrust, but we definitely should have one. Could you please file a documentation issue, so we do not forget to write developer overview?

[jeromeku]

@gevtushenko

Thanks!

Filed here #1119.