Add single precision option to the dycore (intermediate state)

.cscs-ci/all.yml

@@ -13,6 +13,7 @@ variables:
   LEVELS: "all"
   GRIDS: "all"
   TOOLS_SUBSETS: "all"
+  FLOAT_PRECISIONS: "all"
 
 
 build_baseimage_aarch64:

.cscs-ci/base.yml

@@ -130,6 +130,7 @@ variables:
     ICON4PY_CI_GT4PY_BUILD_CACHE_BASE_DIR: /gt4py_cache
     ICON4PY_DALLCLOSE_PRINT_INSTEAD_OF_FAIL: false
     ICON4PY_DRIVER_LOGGING_LEVEL: critical
+    ICON4PY_FLOAT_PRECISION: "${FLOAT_PRECISION}"
     PYTEST_ADDOPTS: "--durations=0"
     # gt4py skips formatting generated code when `$CLANG_FORMAT_EXECUTABLE --version` fails.
     CLANG_FORMAT_EXECUTABLE: /bin/false

.cscs-ci/benchmark_bencher_common.yml

@@ -2,6 +2,7 @@
   variables:
     ICON4PY_ENABLE_GRID_DOWNLOAD: true
     ICON4PY_ENABLE_TESTDATA_DOWNLOAD: true
+    FLOAT_PRECISION: "double"
     PYTHONOPTIMIZE: 2
     STENCIL_TEST_SELECTION: "test_Test and compile_time_domain"
     GRANULE_TEST_SELECTION: "test_diffusion_benchmark or test_benchmark_solve_nonhydro"

.cscs-ci/default.yml

@@ -8,7 +8,7 @@ variables:
   # IDs. These are the defaults for the pipeline and don't include all possible
   # jobs. They are meant to be overridden with e.g.
   #
-  # cscs-ci run default;BACKENDS=gtfn_cpu;LEVELS=unit;MODEL_SUBPACKAGES=common:driver;MODEL_MPI_SUBPACKAGES=common;SESSIONS=model;MODEL_SUBSETS=datatest
+  # cscs-ci run default;BACKENDS=gtfn_cpu;LEVELS=unit;MODEL_SUBPACKAGES=common:driver;MODEL_MPI_SUBPACKAGES=common;SESSIONS=model;MODEL_SUBSETS=datatest;FLOAT_PRECISIONS=double
   SESSIONS: "all"
   MODEL_SUBPACKAGES: "all"
   MODEL_SUBSETS: "datatest"
@@ -18,6 +18,7 @@ variables:
   LEVELS: "unit"
   GRIDS: "icon_regional"
   TOOLS_SUBSETS: "unittest"
+  FLOAT_PRECISIONS: "double"
 
 
 build_baseimage_aarch64:

.cscs-ci/merge.yml

@@ -13,6 +13,7 @@ variables:
   LEVELS: "unit:integration"
   GRIDS: "simple:icon_regional"
   TOOLS_SUBSETS: "datatest:unittest"
+  FLOAT_PRECISIONS: "single:double"
 
 .only_merge_queue:
   rules: &only_merge_queue

.github/workflows/mandatory_and_optional_test_reminder.yml

@@ -29,6 +29,7 @@ jobs:
             * `BACKENDS`: backends
             * `GRIDS`: grids for stencil tests (`simple`, `icon_regional`, or `icon_global`)
             * `LEVELS`: testing level for non-stencil tests (`unit` or `integration`)
+            * `FLOAT_PRECISIONS`: `single` or `double`
 
             For each option, `all` can be used as a shorthand for all possible values of that variable, e.g. `LEVELS=all`.
 

AGENTS.md

@@ -72,6 +72,7 @@ If a `shell.nix` exists in the repo root, you can use it to provide these depend
 - `GT4PY_BUILD_CACHE_DIR`: GT4Py stencil compilation cache location
 - `GT4PY_BUILD_JOBS`: limit parallel stencil compilation jobs (unset by default)
 - `PYTEST_ADDOPTS`: default pytest options (xdist workers, verbosity)
+- `ICON4PY_FLOAT_PRECISION`: choose precision setting from `double` (default), `single`, `mixed` (broken)
 
 ### Clean rebuild
 
@@ -119,6 +120,9 @@ uv run --group test --frozen pytest --datatest-skip model/<component>/
 # Only datatests (requires test data):
 uv run --group test --frozen pytest --datatest-only model/<component>/
 
+# Single-precision mode (only runs tests with the pytest marker `single_precision_ready`):
+ICON4PY_FLOAT_PRECISION=single uv run --group test --frozen pytest model/<component>/
+
 # MPI tests (requires mpi4py, distributed extra; always use -n0 for sequential):
 mpirun -np 4 .cscs-ci/scripts/ci-mpi-wrapper.sh uv run --group test --frozen pytest -v -s --with-mpi -n0 -k mpi_tests model/<component>/
 #   --with-mpi: enables MPI test mode (from pytest-mpi plugin)
@@ -139,7 +143,6 @@ Registered by `icon4py.model.testing.pytest_hooks` (auto-loaded via `addopts`):
 | `--datatest-skip`                           | Skip all datatests                                                                |
 | `--backend <name>`                          | GT4Py backend (default: embedded; others: gtfn_cpu, gtfn_gpu, dace_cpu, dace_gpu) |
 | `--grid <name>`                             | Grid to use                                                                       |
-| `--enable-mixed-precision`                  | Switch from double to mixed-precision                                             |
 | `--level {any,unit,integration,validation}` | Filter by `@pytest.mark.level` marker. `any` (default) excludes validation tests. |
 | `--skip-stenciltest-verification`           | Skip verification of StencilTest against reference outputs                        |
 
@@ -182,11 +185,19 @@ Nox mirrors the CI pipeline. Useful for running comprehensive test suites:
 uv run --group test --frozen nox -l
 
 # Run all tests for a specific component and subset:
-uv run --group test --frozen nox -s 'test_common(datatest=True)'
-uv run --group test --frozen nox -s 'test_common(datatest=False)'
+uv run --group test --frozen nox -s "test_model-3.13(datatest, common)"
+
+# Select one subset across all subpackages by tag:
+uv run --group test --frozen nox -t basic
+uv run --group test --frozen nox -t datatest
+
+# Run tests in single-precision mode:
+ICON4PY_FLOAT_PRECISION=single uv run --group test --frozen nox -s "test_model-3.13(basic, dycore)"
 ```
 
-Subset options: `datatest`, `stencils`, `basic` (datatest-skip, no stencils/benchmarks).
+Subset options: `datatest`, `stencils`, `basic` (datatest-skip, no stencils/benchmarks). `test_model_mpi` has no `stencils` subset because stencil tests are serial by definition.
+
+Subpackage IDs are the last path component of the package directory: `tracer_advection`, `diffusion`, `dycore`, `microphysics`, `muphys`, `physics_driver`, `common`, `driver`, `testing`.
 
 ## Triggering CSCS CI
 
@@ -200,4 +211,4 @@ See `.github/workflows/mandatory_and_optional_test_reminder.yml` for the authori
   - `cscs-ci run default;MODEL_SUBPACKAGES=common:driver;SESSIONS=model`
 - The `cscs/merge` pipeline runs automatically on the merge queue; do not trigger it manually. It runs as a dummy pipeline on PR pushes and runs no tests.
 - Some pipelines, especially those running on the normal slrum partition, can in the worst case take hours to schedule (when cluster is busy) and run (see SLURM_TIMELIMIT in the CSCS CI configs). Keep this in mind when waiting for jobs to finish. Test jobs may also need to populate GT4Py caches which can take long.
-- CSCS CI configs are in the ci/ subdirectory. The CI runs using GitLab runners and the configuration is the same as for regular GitLab pipelines.
+- CSCS CI configs are in the `.cscs-ci/` subdirectory. The CI runs using GitLab runners and the configuration is the same as for regular GitLab pipelines.

bindings/src/icon4py/bindings/diffusion_wrapper.py

@@ -163,8 +163,8 @@ def diffusion_init(  # noqa: PLR0917 [too-many-positional-arguments]
     if zd_cellidx is None:
         # then zdiffu_t is False or the list on that rank is empty, then all of the following are not initialized
         assert zd_vertidx is None and zd_intcoef is None and zd_diffcoef is None
-        zd_diffcoef = gtx.zeros(cell_k_domain, dtype=theta_ref_mc.dtype, allocator=allocator)
-        zd_intcoef = gtx.zeros(cell_c2e2c_k_domain, dtype=wgtfac_c.dtype, allocator=allocator)
+        zd_diffcoef = gtx.zeros(cell_k_domain, dtype=wpfloat, allocator=allocator)
+        zd_intcoef = gtx.zeros(cell_c2e2c_k_domain, dtype=wpfloat, allocator=allocator)
         zd_vertoffset = gtx.zeros(cell_c2e2c_k_domain, dtype=xp.int32, allocator=allocator)
     else:
         # transform lists to fields
@@ -183,7 +183,7 @@ def diffusion_init(  # noqa: PLR0917 [too-many-positional-arguments]
                 data_alloc.adjust_fortran_indices(zd_cellidx),
                 data_alloc.adjust_fortran_indices(zd_vertidx),
             ),
-            default_value=gtx.float64(0.0),
+            default_value=wpfloat(0.0),
             allocator=allocator,
         )
         zd_intcoef = data_alloc.scattered_field(
@@ -194,7 +194,7 @@ def diffusion_init(  # noqa: PLR0917 [too-many-positional-arguments]
                 slice(None),
                 data_alloc.adjust_fortran_indices(zd_vertidx),
             ),
-            default_value=gtx.float64(0.0),
+            default_value=wpfloat(0.0),
             allocator=allocator,
         )
         zd_vertoffset = data_alloc.scattered_field(

bindings/tests/bindings/test_diffusion_wrapper.py

@@ -99,11 +99,11 @@ def test_diffusion_wrapper_granule_inputs(  # noqa: PLR0917 [too-many-positional
     exner = test_utils.array_to_array_info(savepoint_diffusion_init.exner().ndarray)
     theta_v = test_utils.array_to_array_info(savepoint_diffusion_init.theta_v().ndarray)
     rho = test_utils.array_to_array_info(savepoint_diffusion_init.rho().ndarray)
-    dtime = savepoint_diffusion_init.get_metadata("dtime")["dtime"]
+    dtime = savepoint_diffusion_init.dtime()
 
     # --- Expected objects that form inputs into init and run functions
     expected_icon_grid = icon_grid
-    expected_dtime = savepoint_diffusion_init.get_metadata("dtime").get("dtime")
+    expected_dtime = savepoint_diffusion_init.dtime()
     expected_edge_geometry: grid_states.EdgeParams = grid_savepoint.construct_edge_geometry()
     expected_cell_geometry: grid_states.CellParams = grid_savepoint.construct_cell_geometry()
     expected_interpolation_state = diffusion_states.DiffusionInterpolationState(
@@ -327,7 +327,7 @@ def test_diffusion_wrapper_single_step(  # noqa: PLR0917 [too-many-positional-ar
     exner = test_utils.array_to_array_info(savepoint_diffusion_init.exner().ndarray)
     theta_v = test_utils.array_to_array_info(savepoint_diffusion_init.theta_v().ndarray)
     rho = test_utils.array_to_array_info(savepoint_diffusion_init.rho().ndarray)
-    dtime = savepoint_diffusion_init.get_metadata("dtime")["dtime"]
+    dtime = savepoint_diffusion_init.dtime()
 
     ffi = cffi.FFI()
     # Call diffusion_init

bindings/tests/bindings/test_dycore_wrapper.py

@@ -268,7 +268,7 @@ def test_dycore_wrapper_granule_inputs(  # noqa: PLR0917 [too-many-positional-ar
     )  # undo the -1 to go back to Fortran value
 
     # other params
-    dtime = sp.get_metadata("dtime").get("dtime")
+    dtime = sp.dtime()
     prepare_fluxes_for_advection = sp.get_metadata("prep_adv").get("prep_adv")
 
     # metric state parameters
@@ -534,7 +534,7 @@ def test_dycore_wrapper_granule_inputs(  # noqa: PLR0917 [too-many-positional-ar
         ),  # TODO(): sp.vol_flx_ic(),
     )
     expected_second_order_divdamp_factor = sp.divdamp_fac_o2()
-    expected_dtime = sp.get_metadata("dtime").get("dtime")
+    expected_dtime = sp.dtime()
     expected_prepare_fluxes_for_advection = sp.get_metadata("prep_adv").get("prep_adv")
     expected_at_first_substep = substep_init == 1
     expected_at_last_substep = substep_init == ndyn_substeps
@@ -809,7 +809,7 @@ def test_granule_solve_nonhydro_single_step_regional(  # noqa: PLR0917 [too-many
     sp_step_exit = savepoint_nonhydro_step_final
 
     # other params
-    dtime = sp.get_metadata("dtime").get("dtime")
+    dtime = sp.dtime()
     prepare_fluxes_for_advection = sp.get_metadata("prep_adv").get("prep_adv")
 
     # solve nh run parameters
@@ -991,7 +991,7 @@ def test_granule_solve_nonhydro_multi_step_regional(  # noqa: PLR0917 [too-many-
     sp_step_exit = savepoint_nonhydro_step_final
 
     # other params
-    dtime = sp.get_metadata("dtime").get("dtime")
+    dtime = sp.dtime()
     prepare_fluxes_for_advection = sp.get_metadata("prep_adv").get("prep_adv")
 
     # solve nh run parameters