This repository contains selected example models for profiling during the hackathon.

• boids_spatial3D: Reynolds boids flocking model with 3D spatial messaging communication.
• boids_rtc_spatial3D: boids_spatial3D, implemented with runtime compiled (RTC) agent functions.
• boids_bruteforce: Reynolds boids flocking model with bruteforce messaging communication.
• boids_rtc_bruteforce: boids_bruteforce, implemented with runtime compiled (RTC) agent functions.

Video of how these models looks is available here.

The dependencies below are required for building FLAME GPU 2 projects.

• CMake >= 3.18
• CUDA >= 11.0 and a Compute Capability >= 3.5 NVIDIA GPU.
  • CUDA >= 10.0 currently works, but support will be dropped in a future release.
• C++17 capable C++ compiler (host), compatible with the installed CUDA version
  • Microsoft Visual Studio 2019 (Windows)
  • make and GCC >= 7 (Linux)
  • Older C++ compilers which support C++14 may currently work, but support will be dropped in a future release.
• git

FLAME GPU 2 uses CMake, as a cross-platform process, for configuring and generating build directives, e.g. Makefile or .vcxproj. This is used to build the FLAMEGPU2 library, examples, tests and documentation.

Below the core commands are provided, for the full guide refer to the main FLAMEGPU2 guide.

These commands can be used to build a high performance version, capable of executing on Marconi for profiling:

mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCUDA_ARCH=70 -DSEATBELTS=OFF -DEXPORT_RTC_SOURCES=ON -DUSE_NVTX=ON
cmake --build . -j `nproc` --target all

FLAMEGPU2 is cross platform, so can be built on Windows too, however you may wish to update CUDA_ARCH=70 if you don't have a volta GPU.

mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCUDA_ARCH=70 -DSEATBELTS=OFF -DUSE_NVTX=ON -DEXPORT_RTC_SOURCES=ON -A x64
ALL_BUILD.sln

The relevant version of the main FLAMEGPU2 repository will be downloaded via CMake/git at configure time, into <build>/_deps/FLAMEGPU2-src. Instead, you can provide your own local copy of FLAMEGPU2, via FETCHCONTENT_SOURCE_DIR_FLAMEGPU2. This will only re-use source files, and not re-use any build directories on disk at the other local location.

i.e.

cmake .. -DFETCHCONTENT_SOURCE_DIR_FLAMEGPU2=/path/to/FLAMEGPU2

The below commands will execute the respective models for several steps, and generate both a timeline and full Nsight Compute data collection.

nsys profile -o boids_timeline ./bin/linux-x64/Release/boids_spatial3D -s 10
ncu --set full -o boids_ncu ./bin/linux-x64/Release/boids_spatial3D -s 5

Note: On first run RTC agent functions will be compiled at runtime, this may take upto 2 minutes to complete. Further runs will pull the precompiled agent functions from a cache. RTC cache hits will fail if the body of the agent function, the dynamic RTC header or the underlying FLAMEGPU2 lib commit hash has changed.

nsys profile -o boids_rtc_timeline ./bin/linux-x64/Release/boids_rtc_spatial3D -s 10
ncu --set full -o boids_rtc_ncu ./bin/linux-x64/Release/boids_rtc_spatial3D -s 5

The number of agents maps cleanly to the number of active threads, by default these models have 32768 agents each.

In either main.cu file you can find the line env.newProperty("POPULATION_TO_GENERATE", 32768u);. This value controls the number of agents and can be updated to test different scales.

FLAMEGPU1 versions of this model for comparison can be found in this branch of the FLAMEGPU repo: https://github.com/FLAMEGPU/FLAMEGPU/tree/boids-performance