cudf_column's Introduction

Hello, I'm jrhemstad 👋

About Me

🔭 I'm currently the Engineering Manager for the team that maintains the CUDA C++ Core Libraries at NVIDIA.
🌱 Previously, I was deeply involved in the RAPIDS ecosystem, where I was a primary maintainer and designer for the libcudf and RMM libraries.
🚀 I helped kickstart and maintain the cuCollections project for GPU-acelerated data structures like hash tables.
💡 I'm passionate about building high-quality, high-performance CUDA C++ software.
🎯 My mission is to make CUDA C++ more delightful for developers.

Publications & Presentations

Revamping llm.c with the CUDA C++ Core Libraries
GTC 2024: Mastering CUDA C++: Best Practices with the CUDA C++ Core Libraries
Using CUDA Stream-Ordered Memory Allocator (Part 1)
Using CUDA Stream-Ordered Memory Allocator (Part 2)
Maximizing Performance with Massively Parallel Hash Maps on GPUs
GTC 2019: RAPIDS CUDA DataFrame Internals for C++ Developers

Connect with Me

📄 LinkedIn

cudf_column's People

Contributors

Watchers

cudf_column's Issues

Allocator Design

Need an allocator design to allow controlling how column's device memory is allocated.

Initial thoughts are to use a polymorphic (non-template) allocator, based on std::pmr::memory_resource and Thrust's thrust::mr::memory_resource.

`device_buffer` and `host_buffer` classes

Need device_buffer and host_buffer classes that are more trivial versions of thrust::device_vector and std::vector.

device_buffer will be the object in which a column's data and bitmask will be stored.

column_view

Design a column_view class as a non-owning reference to the contents of a column. Copy constructor and destructor are trivial.

Furthermore, consider making all libcudf APIs take column_views as inputs instead of columns. Ideally, a column_view can be passed by value straight into a kernel without detriment.

Research std::string_view design for inspiration.

Recommend Projects