Testing

Cocoa uses Catch2 for the unit tests, with Kokkos/Trilinos initialization handled in a custom main entry point. The integration tests are ctest entries that run the cocoa binary and compare its output against a committed reference with a Python script.

Test Framework

Tests are written with Catch2 (v3) TEST_CASE and SECTION macros:

#include <catch2/catch_test_macros.hpp>

TEST_CASE("Mesh", "[mesh]") {
  // Setup code runs for each section
  const auto mesh = create_test_mesh();

  SECTION("Valid mesh construction") {
    REQUIRE(mesh.num_nodes() == 9);
    REQUIRE(mesh.num_elements() == 8);
  }

  SECTION("Coordinate access") {
    const auto& nodes = mesh.nodes();
    REQUIRE_THAT(nodes[0].location().x(), WithinAbs(-88.0, 1e-10));
  }
}

The test executable uses a custom main() to initialize Tpetra/Kokkos before running tests:

#define CATCH_CONFIG_RUNNER
#include <catch2/catch_all.hpp>
#include "Tpetra_Core.hpp"

int main(int argc, char* argv[]) {
  Tpetra::ScopeGuard tpetra_scope_guard(&argc, &argv);
  return Catch::Session().run(argc, argv);
}

Building and Running Tests

# Build with tests (enabled by default)
cmake --preset dev
cmake --build --preset dev

# Run all tests via CTest, from anywhere in the source tree
ctest --preset dev

# Run the test executable directly (paths below are relative to build/)
cd build
./test/cocoa_tests

# Run specific test cases by tag
./test/cocoa_tests "[mesh]"
./test/cocoa_tests "[gwce]"

# Run specific test case by name
./test/cocoa_tests "Mesh"

# List all available tests
./test/cocoa_tests --list-tests

Test Files

Test sources live in test/ as test_<component>.cpp, one file per component, with test_main.cpp as the entry point. The current set is COCOA_TEST_SOURCES in test/CMakeLists.txt. ./test/cocoa_tests --list-tests enumerates the test cases and --list-tags the tags they are filtered by.

Test Data

Test data and helper utilities are organized as follows:

test/
├── TestMeshData.hpp              # Test mesh generation utilities
├── TestPrecision.hpp             # Mixed-precision test tolerance helpers
├── test_*.cpp                    # Test source files
├── run_in_workdir.cmake          # Runs a driver in a private fixture copy
├── run_integration_test.cmake    # Default integration test driver
├── run_*_test.cmake              # Other drivers (checkpoint, serial-vs-MPI, ...)
├── regenerate_references.sh      # Reference regeneration (see below)
└── data/
    ├── test_config.yaml          # ConfigurationReader test fixture
    ├── baroclinic/               # Baroclinic forcing fixtures
    ├── channel/                  # Channel and beach cases (+ validation)
    ├── flow_channel/             # Flow-boundary case
    ├── global_case/              # Global tidal case
    ├── ice_drag/                 # Ice drag golden values
    ├── internal_tide_checkpoint/ # Internal tide checkpoint case
    ├── internal_weir/            # Internal weir case
    ├── meteo/                    # Per-format meteorological reader fixtures
    ├── partition/                # Partition cache fixtures
    ├── vortex/                   # ATCF tracks and GAHM control solutions
    └── wnat/                     # WNAT integration test data
        ├── wnat.nc                             # Mesh file
        ├── cocoa_config_*.yaml                 # One config per case
        ├── compare_tides_reference.py          # Reference comparison script
        ├── plot_comparison.py                  # Comparison plot generator
        └── *_reference{,_mpi}.nc               # Serial and MPI references

Runnable example cases live in the cocoa-examples repository; the fixtures above are separate copies maintained with the tests.

See Test Cases for detailed descriptions of each test case, including expected results and verification plots.

Two directories hold unit-test fixtures rather than integration ones:

• ice_drag/lupkes_golden.csv: generate_lupkes_golden.py computes the Lupkes drag blend in Python from the paper formulas, and test_ice_drag.cpp holds the C++ implementation to it at 1e-12.

• baroclinic/: a 25-node tiny_mesh.nc and a 3-snapshot nodal_forcing.nc with analytic values (make_test_fixture.py), malformed variants of that file for the reader’s error paths, and ADCIRC-derived control tables for the internal-tide drag and dispersion-friction unit tests under reference/.

The TestMeshData.hpp helper provides functions to create small test meshes:

#include "TestMeshData.hpp"

using TestData::TestMeshData;

TEST_CASE("Example", "[example]") {
  const auto [x_coords, y_coords] = TestMeshData::create_coordinates();
  const auto elevation = TestMeshData::create_elevation();
  const auto triangles = TestMeshData::create_triangles();
  const auto boundaries = TestMeshData::empty_boundaries();

  // Construct HostMesh (held by unique_ptr), then wrap in Mesh
  auto host_mesh = std::make_unique<Geometry::HostMesh>(
      x_coords, y_coords, elevation, triangles, boundaries,
      TestMeshData::origin, TestMeshData::projection_type);
  const Cocoa::Mesh mesh(std::move(host_mesh));
}

Catch2 Assertions

Common assertions used in Cocoa tests:

Basic Assertions:

REQUIRE(expr);           // Fatal assertion
CHECK(expr);             // Non-fatal assertion
REQUIRE_FALSE(expr);     // Require expression is false
REQUIRE_NOTHROW(expr);   // Require no exception thrown
REQUIRE_THROWS_AS(expr, ExceptionType);  // Require specific exception

Floating-Point Comparisons:

#include <catch2/catch_approx.hpp>
#include <catch2/matchers/catch_matchers_floating_point.hpp>

using Catch::Approx;
using Catch::Matchers::WithinAbs;
using Catch::Matchers::WithinRel;

REQUIRE(value == Approx(expected));              // Default tolerance
REQUIRE(value == Approx(expected).epsilon(1e-6)); // Custom tolerance
REQUIRE_THAT(value, WithinAbs(expected, 1e-10)); // Absolute tolerance
REQUIRE_THAT(value, WithinRel(expected, 1e-6)); // Relative tolerance

Writing New Tests

1. Create or edit a test file in test/:

#include <catch2/catch_test_macros.hpp>
#include "cocoa_kernel/geometry/Mesh.hpp"

TEST_CASE("ComponentName", "[component][tag]") {
  SECTION("Descriptive behavior name") {
    // Arrange
    auto obj = create_object();

    // Act
    auto result = obj.method();

    // Assert
    REQUIRE(result == expected);
  }
}

2. Add the file to test/CMakeLists.txt if creating a new file:

set(COCOA_TEST_SOURCES
    test_main.cpp
    # ... existing files ...
    test_your_new_file.cpp)

3. Rebuild and run:

cmake --build --preset dev --target cocoa_tests
./build/test/cocoa_tests "[your_tag]"

Compile-Time Tests

For constexpr functionality, use static_assert:

// Compile-time tests for constexpr functions
static_assert(TimeDelta::fromDays(1).totalHours() == 24,
              "Day to hours conversion failed");
static_assert(TimeDelta::fromHours(24) == TimeDelta::fromDays(1),
              "Equality comparison failed");

Integration Tests

The integration tests run complete simulations, most on the Western North Atlantic (WNAT) mesh, and check that the full pipeline (mesh loading, tidal forcing with SAL, time stepping, output) reproduces committed reference results within a per-test tolerance. Output always goes through the background writer thread, so every run exercises the asynchronous write path.

Integration tests span several categories: WNAT tides (explicit and implicit GWCE, serial and MPI), the GAHM vortex-only and composed-with-gridded-met cases, the global tidal case (serial, split-output, and MPI), the internal weir case, and checkpoint/restart round-trips. Each test runs in a private working directory (see Integration Working Directories below). The list of integration tests, with their names, rank counts, and configs, is test/CMakeLists.txt. Search it for NAME Integration_ to enumerate them, or run ctest -N -L integration.

The WNAT tide tests run a 1-day simulation (M2 boundary + tide potential + SAL) and compare the output against a serial reference solution: the explicit case at \(\Delta t\) = 20 s, with CF meteorological forcing also read, and the implicit case at \(\Delta t\) = 60 s. The MPI variants use a separate reference because mesh partitioning changes floating-point operation order.

Integration test tolerances are NOT uniform: they range from 1e-3 to 1e-12 depending on what the comparison must absorb. Runs that must agree with no arithmetic difference use the tightest tolerances (1e-12 for two configurations that must produce the same run, 1e-10 for checkpoint round-trips with no intervening solver iteration); 1e-6 covers most serial simulation comparisons. MPI runs are relaxed to 1e-3 because partitioning changes the order of floating-point reductions; checkpoint restart through the iterative consistent-GWCE solver is relaxed to 1e-5 because the CG solver’s own convergence tolerance is 1e-5. Each -DTOLERANCE= value in test/CMakeLists.txt carries a comment explaining why a tighter value would fail. Those comments take precedence over this page; update a tolerance’s comment whenever you change the tolerance.

The comparison script also verifies that wet/dry state (NaN locations) matches between test and reference at every timestep. The comparison scripts (compare_bitwise.py, compare_tides_reference.py) live in test/data/wnat/ and are shared across test areas.

The fixtures under test/data/ pin physics.wetting_velocity: inertial, the form Cocoa used for every run before 0.6.0, so their committed references carry that wetting rule; the default is linear.

After comparison, a Cartopy-based plotting script generates a visual summary (wnat_comparison_serial.png, wnat_comparison_mpi.png) with peak water level and velocity maps, difference maps, and time series at probe points. Plots are kept in build/test/integration_plots; in CI, they are uploaded as a GitHub Actions artifact.

Prerequisites:

Integration tests require Python 3 with numpy, xarray and netCDF4; cartopy is needed only by the comparison-plot script. If an import fails, CMake reports Integration tests disabled at configure time and skips them.

CMake options:

Integration tests are off by default:

# Enable integration tests
cmake --preset dev -Dcocoa_ENABLE_INTEGRATION_TESTS=ON

The dev-maintainer, ci and ci-coverage presets already enable them.

In CI, integration tests are enabled for the serial sanitizer job (build-and-test) and the coverage job. The GPU compile-check job (gpu-build) does not run tests.

Running integration tests selectively:

# Run only integration tests
ctest -R Integration

# Run only the serial integration test
ctest -R Integration_WNAT_tides$

# Run only the MPI integration test
ctest -R Integration_WNAT_tides_mpi

# Exclude integration tests (run unit tests only)
ctest -E Integration

Regenerating reference solutions:

If numerical methods change and produce different (but correct) results, the reference solutions must be regenerated with test/regenerate_references.sh. Do not hand-run cocoa and mv the output: the script defines which config produces which reference file, and it enforces the conditions those references depend on:

# Regenerate everything (wnat, global, weir, gahm, baroclinic, channel)
test/regenerate_references.sh --cocoa build/src/cocoa/cocoa

# Regenerate one case only
test/regenerate_references.sh --cocoa build/src/cocoa/cocoa wnat

Run it inside the build container (or an environment with the same compiler, Trilinos, and Kokkos as CI) so the regenerated references are reproducible: every run is single-threaded (OMP_NUM_THREADS=1), so the bytes it produces are bit-reproducible run to run on the same toolchain. The script clears any stale cocoa_output.nc before each run so a failed run cannot leave a misleading reference, and it discards the transient cocoa_output_peak.nc file rather than storing it. It regenerates every reference listed in its header comment (the WNAT, GAHM, moving-ice, baroclinic, global, internal-weir and channel references, with 2-rank MPI references where a case has one); the split-output variants and the weir MPI variant compare against the serial reference rather than carrying their own, so those are not separately regenerated.

Commit the regenerated .nc files with provenance in the commit body: the commit and config that produced them, and why (what numerical change made the old references stale). Never regenerate references to fix an unrelated failing test: a reference that moves without a documented reason can no longer catch a real regression.

Working directories:

Each integration test runs in its own copy of test/data, build/test/integration/<test name>, with its case at <test name>/<case>, so tests never share output files and ctest -j runs them side by side. The copy leaves out the files a run writes (cocoa_output*, cocoa_checkpoint*, plots); a fixture’s ../channel/... paths resolve inside it, and nothing a run writes reaches the source tree. A passing test deletes its directory and keeps its plots in build/test/integration_plots; a failing one leaves the directory for inspection. Register new end-to-end tests with cocoa_add_integration_test (cmake/CocoaIntegrationTests.cmake); a file a fixture test generates reaches the consumer through its EXTRA_INPUTS.

Integration Test Fixtures

The tests below run on the in-repository cases described in Test Cases. Entries with an _mpi suffix are registered only when MPI tests are enabled.

Ideal channel

Fixtures in test/data/channel/ (Ideal Channel (Quadrilateral Validation)):

• Integration_channel_straight_tide, Integration_channel_meander_flow and their _implicit variants: the coarse hybrid meshes against committed references at 1e-6 on zeta, ucx and ucy. Each has an _mpi variant that compares a 2-rank run against a 2-rank reference at 1e-3.

• Integration_channel_straight_flow: the uniform-flow smoke case (completes, finite, bounded, hybrid encoding checked), lumped only.

• Integration_hybrid_tide_lumped and Integration_hybrid_tide_consistent: an M2 smoke run per mass matrix on channel_hybrid_tide.nc; Integration_hybrid_tide_mpi_lumped and _consistent hold a 2-rank run to the serial one.

• Integration_beach_cut_wetdry_mpi: the coarse triangle beach on 2 ranks through a committed partition cache (channel_beach_coarse_triangle_tide.nc.partition_2.nc, written by make_beach_cut_partition.py) whose cut the tidal front crosses. Every output’s wet/dry set must equal the serial run’s.

• Integration_broad_inlet_flow_mpi2/3/4 and Integration_two_inlets_flow_mpi2/3/4: wide specified-flow inlets the partitioner has to cut, held to the serial run on the fields and on the cross-section discharge.

• Integration_surplus_tide_values_rejected and its _mpi variant: a tidal boundary_values list one entry longer than the open boundary must be rejected the same way at every rank count.

The hybrid tests check the output element encoding (nvertex of 4, -1 in each triangle row’s fourth column) with check_hybrid_element_encoding.py. Two Fixture_channel_* entries check the fixtures themselves: a cocoa-plot render of the hybrid strip’s mesh, and an exact-match check of the committed meshes against the generator. To regenerate the meshes and run that check by hand:

cd test/data/channel
python make_channel_meshes.py .
python make_channel_meshes.py --check-committed .

Internal weir

Fixtures in test/data/internal_weir/ (Internal Weir):

• Integration_internal_weir: compares zeta at every output time against weir_reference.nc at 1e-6; the wet/dry pattern must match. Integration_internal_weir_mpi holds a 2-rank run to the same reference at 1e-3.

• Integration_weir_checkpoint: runs the full day, then a cold start to 12:00 that writes a checkpoint, then a restart from it to the end. The restart’s zeta must match the full run at shared output times to 1e-10. Integration_weir_checkpoint_mpi runs it on 2 ranks at 1e-3.

• Integration_internal_tide_checkpoint_mid_window (in test/data/internal_tide_checkpoint/): the same checkpoint test with internal-tide wave drag active (Baroclinic Coupling). The mesh internal_weir_it.nc is the weir mesh plus a 3-column (Lyard) internal_tide_friction attribute. The forcing file internal_tide_baroclinic_forcing.nc is a single frozen snapshot with n_bottom = 0.02, n_mean = 0.01 and zero BPG. The cold start ends at 11:55:00 (42900 s), which is not a multiple of the 720 s de-tiding sampling window, so the restart must resume the de-tiding ring mid-window. Its configs add physics.internal_tide_friction: mesh and a forcing.baroclinic block with internal_tide.enabled: true. Tolerance 1e-10; the _mpi variant runs on 2 ranks at 1e-3.

GAHM vortex

Fixtures in test/data/wnat/ and test/data/vortex/ (GAHM Parametric Vortex (Katrina 2005)). Each test compares zeta against a committed reference at 1e-6 with compare_tides_reference.py.

• Integration_gahm_vortex_only: the vortex alone, with no blend ring, against gahm_vortex_only_reference.nc.

• Integration_gahm_composed: the vortex plus the out-of-window gridded reader, against gahm_composed_reference.nc. It guards that the vortex does not vanish outside the reader’s time window.

• Integration_gahm_composed_domains: the composed config with the gridded reader moved into a one-entry domains: list. It runs against the same reference as the flat form, so a domains: singleton must be indistinguishable from the flat keys.

Sea ice

Fixtures in test/data/wnat/ (Sea Ice (WNAT)):

• Integration_wnat_zero_ice: runs the no-ice config and the zero-ice config with the same binary. The primary check requires the two outputs to be identical (tolerance 0.0). A secondary check compares the zero-ice run against wnat_reference.nc at 1e-6. The zero-ice config turns off ice_concentration output so the two variable sets match. The Fixture_wnat_zero_ice_source entry generates its met file.

• Integration_wnat_moving_ice: compare_ice_moving.py checks that zeta differs from a no-ice run by more than 1e-4, that ice_concentration matches an independent bilinear interpolation of wnat_moving_ice.nc at every output time to 1e-6, and that the run matches wnat_moving_ice_reference.nc at 1e-6.

Baroclinic forcing

Fixtures in test/data/wnat/ (Baroclinic Forcing (WNAT)). Both tests run only on host backends: they require run-to-run determinism, and atomic scatter on GPU backends is not deterministic.

• Integration_WNAT_baroclinic_zero_forcing: runs the implicit config with baroclinic off, then with baroclinic on against the all-zero file. compare_bitwise.py requires the outputs to be identical (no tolerance) and the extra bpg_effective_x/bpg_effective_y outputs to be exactly zero.

• Integration_WNAT_baroclinic_composed: runs cocoa_config_baroclinic_composed_off.yaml and _on.yaml. The ON run must differ from the OFF run by more than 0.05 m in zeta at the final output time, or in its wet/dry set. It must also match baroclinic_composed_reference.nc at 1e-6.

Python Tools Tests

The cocoa-tools package (utils/cocoa_tools) has its own pytest suite, which is not part of ctest. Install the package and run the suite from its directory:

pip install ./utils/cocoa_tools pytest
cd utils/cocoa_tools
python -m pytest -q

CI runs the same suite in the cocoa_tools job. The mesh-tools tests read fixtures from test/data.

Debugging Failed Tests

# Run with verbose output
./test/cocoa_tests -s "[failing_tag]"

# Run under debugger
lldb ./test/cocoa_tests -- "[failing_tag]"

# Run specific named test
./test/cocoa_tests "Exact Test Name"

# Show test durations
./test/cocoa_tests -d yes