This is a repo of simple examples from Mark Harris's presentation for the 2023 GPU Technology Conference, entitled: "Robust and Efficient CUDA C++ Concurrency with Stream-Ordered Allocation", or "Stream Safety First!".
The examples are provided in the form of Catch2 unit tests where each test has one or more sections demonstrating some unsafe concurrent CUDA C++ code pattern, and then one or more sections demonstrate a safe alternative.
- NVIDIA CUDA Toolkit. Developed on versions 11.5 and 11.8. A version at least as new is recommended.
- CMake 3.20.1 or newer
Only tested on Linux (Ubuntu 20.04).
The CMake configuration uses the CMake Package Manager, CPM to fetch all other dependencies (Catch2 and RAPIDS Memory Manager). To build, simply run cmake in an out-of-source build directory, and then run make.
Example:
cd stream-safety-first
mkdir build && cd build
cmake ..
make
To use Ninja instead of Make:
cd stream-safety-first
mkdir build && cd build
cmake .. -GNinja
ninja
Running the built executable stream-safety-tests will run all unit tests. Note that multiple tests
exercise undefined behavior so it is possible that they may fail.
./stream-safety-tests
===============================================================================
All tests passed (17 assertions in 7 test cases)