Skip to content

Initial benchmark#11

Draft
ptheywood wants to merge 15 commits into
mainfrom
initial-benchmark
Draft

Initial benchmark#11
ptheywood wants to merge 15 commits into
mainfrom
initial-benchmark

Conversation

@ptheywood

@ptheywood ptheywood commented May 11, 2026

Copy link
Copy Markdown
Collaborator

Adds the initial repostiory structure, (set of) benchmark models and postprocessing & documentation.

  • Replaced template readme content with the actual content
  • Adds the FP32 version of the circles benchmark model, with a sweep over population size with a fixed initial density
  • Adds the FP64 version of the circles benchmark model, with a sweep over population size with a fixed initial density

Closes #2
Closes #5

Note

This currently fetches the amdgpu branch from the flamegpu/FLAMEGPU2 repository which (at the time of writing) has not been merged into the master branch.
Once this merge occurs, the fetched version can be updated to require v2.0.0-rc.5 or newer (i.e. the next release candidate), and any prior commits with the amdgpu reference may stop working.

@ptheywood ptheywood force-pushed the initial-benchmark branch from 8cb2aa3 to 20fb978 Compare May 20, 2026 15:53
ptheywood and others added 15 commits June 11, 2026 14:06
- Uses a repository structure based on FLAMEGPU/flamegpu2-model-template-cpp
- CMake project which fetched the required verison of FLAME GPU
- Single binary file which runs a sweep of multiple benchmarks
- Emits JSON including build and runtime metadata + benchmark data
- Includes a python script for plotting benchmark data
- Repetitions of a given problem size use slightly offset seeds
This is not a fully FP64 benchmark due to Spatial3D messaging only being fp32 currently, although this can be imprvoed in the future

It is sufficient to highlight the difference between consumer and hpc gpus.

Cannot soley use a typedef/templating for this, due to presence of sinf/sin etc. Can be improved in the future to reduce duplication
CUDA 12.4 on TUoS stanage is unhappy about this even with CUDAHOSTCXX specifying a newer host compiler
…nd median estimator

- Improve the plotting script to handle more than a single model
- Optional SVG output + increase PNG DPI
- Change plotting estimator to median, due to err_style='bars' issue with large outliers?
@ptheywood ptheywood force-pushed the initial-benchmark branch from e360640 to 5a4bce2 Compare June 11, 2026 13:17
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Benchmark: FP64 Initial FLAME GPU benchmark

1 participant