Skip to content

GPU Coagulation Validation Record

This maintainer and release-review record documents evidence for the existing direct GPU-coagulation path. It does not introduce a production API, physics, public export, runnable integration, CPU fallback, graph capture/replay, autodiff, adaptive stepping, or a performance claim.

The fixed-mask matrix has exactly 11 executable rows. Input validation is deterministic validation in particula/gpu/kernels/tests/coagulation_validation_test.py, backed by the Warp-free independent support equations in particula/gpu/kernels/tests/_coagulation_validation_support.py. numerical validation public- step invariant and ownership assertions are in particula/gpu/kernels/tests/coagulation_validation_test.py, supported by particula/gpu/kernels/tests/_coagulation_public_step_support.py. integration validation is the bounded stochastic evidence in particula/gpu/kernels/tests/coagulation_stochastic_validation_test.py.

Mask Canonical mechanism tuple input validation deterministic numerical validation invariants/ownership integration validation stochastic
1 ("brownian",) input validation numerical validation integration validation
2 ("charged_hard_sphere",) input validation numerical validation integration validation
4 ("sedimentation_sp2016",) input validation numerical validation integration validation
8 ("turbulent_shear_st1956",) input validation numerical validation integration validation
3 ("brownian", "charged_hard_sphere") input validation numerical validation integration validation
5 ("brownian", "sedimentation_sp2016") input validation numerical validation integration validation
6 ("charged_hard_sphere", "sedimentation_sp2016") input validation numerical validation integration validation
9 ("brownian", "turbulent_shear_st1956") input validation numerical validation integration validation
10 ("charged_hard_sphere", "turbulent_shear_st1956") input validation numerical validation integration validation
12 ("sedimentation_sp2016", "turbulent_shear_st1956") input validation numerical validation integration validation
15 ("brownian", "charged_hard_sphere", "sedimentation_sp2016", "turbulent_shear_st1956") input validation numerical validation integration validation

The three-way masks 7, 11, 13, and 14 are deferred and fail closed. They are not supported combinations, and this record makes no claim for any other combination.

  • input validation deterministic: Brownian pair-rate comparisons use rtol=1e-7, atol=0. Brownian property and selector-majorant checks, along with other applicable positive or additive rate, property, and majorant comparisons, use rtol=1e-6, atol=0. Physical zeros are exact.
  • numerical validation invariants and ownership: concentration-weighted, per-box, per-species inventory uses rtol=1e-12, atol=1e-30. The same public-step evidence covers applicable charge conservation, legal pairs, merge bookkeeping, inactive slots, caller-buffer identity, persistent RNG lifecycle, and atomic preflight failures.
  • integration validation stochastic: each executable row uses one-proposal capacity, 100 unique fresh seeds, an independent initial-state expectation, and the 3 * sqrt(expected_mean) bound. integration validation is not conservation evidence and is not an exact accepted-pair, seed, RNG, CPU/Warp, or CUDA replay guarantee.

Warp CPU is the required baseline when Warp is installed. CUDA is optional, local/manual additive evidence and skips cleanly when unavailable. The warp, cuda, gpu_parity, and stochastic markers describe tests; selecting markers does not select a device. numerical validation enumerates available Warp devices, so it includes available CUDA even when numerical validation tests are not cuda-marked.

Terminal window
pytest particula/gpu/kernels/tests/coagulation_validation_test.py -q -m "warp and gpu_parity" -Werror
pytest particula/gpu/kernels/tests/coagulation_stochastic_validation_test.py -q -m "warp and stochastic and not cuda" -Werror
pytest particula/gpu/kernels/tests/coagulation_stochastic_validation_test.py -q -m "warp and cuda" -Werror
pytest particula/gpu/kernels/tests/coagulation_test.py -q -Werror

Missing Warp or unavailable CUDA is valid only when an existing guarded test skips. A skip is not CPU/Warp or CUDA parity evidence.

Recorded coagulation benchmark evidence describes a dated, machine-specific tradeoff. It is separate from the physics validation above and from current resident/capture qualification.

Explore this topic: Scientific and numerical correctness