gpu-cuda.md

GPU tracing

SIMPLE's supported production GPU path is native CUDA. Configure it in the config namelist in simple.in; no environment variable selects the backend or integrator.

gpu_mode = 'production'
gpu_backend = 'cuda-native'
gpu_method = 'dopri'
gpu_landreman = .True.       ! segmented loss objective; see arXiv:2606.19523

Use Boozer guiding-center tracing (isw_field_type = 2) without collisions, walls, orbit output, or classifier options. Native CUDA supports tuned DOPRI and Cash-Karp. DOPRI is the default. Both use relerr as their tolerance. The segmented loss-threshold objective is the Landreman comparison mode; the related optimization and tracing protocol is described in https://arxiv.org/abs/2606.19523.

Build and run

Requirements are the NVIDIA CUDA toolkit, a C++17 compiler, GNU or NVIDIA Fortran, and the normal SIMPLE dependencies.

cmake -S . -B build-cuda -G Ninja -DSIMPLE_ENABLE_CUDA=ON
cmake --build build-cuda -j
(cd run && ../build-cuda/simple.x simple.in)

gpu_compact_init = .True. enables the compact Boozer initializer. Use it only with native CUDA and Boozer starts. gpu_start_coordinates = 'boozer' means that start.dat already contains Boozer coordinates.

Production controls

The defaults are:

gpu_num_devices = 1
gpu_t_block = 1d-3
gpu_loss_tau = 0.1d0
gpu_maxloss = 0.02d0
gpu_min_timestep = 1d-10
gpu_pilot_fraction = 0.04d0
gpu_work_order = .True.
gpu_particle_profile = ''

gpu_particle_profile writes initial and final states, field-evaluation counts, loss steps, and loss times. Multiple devices are experimental; use one process per GPU unless device-local field tables have been arranged.

The native CUDA tests are test_cuda_native_rk54 and test_cuda_native_bridge; the latter checks segmented loss accounting through the Fortran-to-CUDA production bridge.