[Experimental] Debugger

ndsl/debug/__init__.py

@@ -0,0 +1,4 @@
+from .config import ndsl_debugger
+
+
+__all__ = ["ndsl_debugger"]

ndsl/debug/config.py

@@ -0,0 +1,51 @@
+"""
+This module provides configuration for the global debugger `ndsl_debugger`
+
+When loading, the configuration will be searched in the global environment variable
+`NDSL_DEBUG_CONFIG`
+
+Configuration is a yaml file of the shape
+```yaml
+stencils_or_class:
+  - copy_corners_x_nord
+  - copy_corners_y_nord
+  - DGridShallowWaterLagrangianDynamics.__call__
+track_parameter_by_name:
+  - fy
+```
+
+Global variable:
+    ndsl_debugger: Debugger accessible throughout the middleware, default to `None`
+        if there is no configuration
+"""
+
+import os
+
+import yaml
+
+from ndsl.comm.mpi import MPIComm
+from ndsl.debug.debugger import Debugger
+from ndsl.logging import ndsl_log
+
+
+ndsl_debugger = None
+
+
+def _set_debugger():
+    config = os.getenv("NDSL_DEBUG_CONFIG", "")
+    if not os.path.exists(config):
+        if config != "":
+            ndsl_log.warning(
+                f"NDSL_DEBUG_CONFIG set but path {config} does not exists."
+            )
+        else:
+            return
+    with open(config) as file:
+        config_dict = yaml.load(file.read(), Loader=yaml.SafeLoader)
+    global ndsl_debugger
+    ndsl_debugger = Debugger(rank=MPIComm().Get_rank(), **config_dict)
+    ndsl_log.info("[NDSL Debugger] On")
+    ndsl_log.debug(f"[NDSL Debugger] Config:\n{config_dict}")
+
+
+_set_debugger()

ndsl/debug/debugger.py

@@ -0,0 +1,109 @@
+import dataclasses
+import numbers
+import os
+import pathlib
+
+import pandas as pd
+import xarray as xr
+
+from ndsl.logging import ndsl_log
+from ndsl.quantity import Quantity
+
+
+@dataclasses.dataclass
+class Debugger:
+    """Debugger relying on `ndsl.debug.config` for setup capable
+    of doing automatic data save on external configuration."""
+
+    # Configuration
+    stencils_or_class: list[str] = dataclasses.field(default_factory=list)
+    track_parameter_by_name: list[str] = dataclasses.field(default_factory=list)
+    save_compute_domain_only: bool = False
+    dir_name: str = "./"
+
+    # Runtime data
+    rank: int = -1
+    calls_count: dict[str, int] = dataclasses.field(default_factory=dict)
+    track_parameter_count: dict[str, int] = dataclasses.field(default_factory=dict)
+
+    def _to_xarray(self, data, name) -> xr.DataArray:
+        if isinstance(data, Quantity):
+            if self.save_compute_domain_only:
+                mem = data.field
+                shp = data.field.shape
+            else:
+                mem = data.data
+                shp = data.shape
+        elif hasattr(data, "shape"):
+            mem = data
+            shp = data.shape
+        elif (
+            pd.api.types.is_numeric_dtype(data)
+            or pd.api.types.is_string_dtype(data)
+            or isinstance(data, numbers.Number)
+        ):
+            return xr.DataArray(data)
+        else:
+            ndsl_log.error(f"[Debugger] Cannot save data of type {type(data)}")
+            return xr.DataArray([0])
+        return xr.DataArray(mem, dims=[f"dim_{i}_{s}" for i, s in enumerate(shp)])
+
+    def track_data(self, data_as_dict, source_as_name, is_in) -> None:
+        for name, data in data_as_dict.items():
+            if name not in self.track_parameter_by_name:
+                continue
+
+            if name not in self.track_parameter_count:
+                self.track_parameter_count[name] = 0
+            count = self.track_parameter_count[name]
+
+            path = pathlib.Path(f"{self.dir_name}/debug/tracks/{name}/R{self.rank}/")
+            os.makedirs(path, exist_ok=True)
+            path = pathlib.Path(
+                f"{path}/{count}_{name}_{source_as_name}-{'In' if is_in else 'Out'}.nc4"
+            )
+            try:
+                self._to_xarray(data, name).to_netcdf(path)
+            except ValueError as e:
+                from ndsl import ndsl_log
+
+                ndsl_log.error(f"[Debugger] Failure to save {data}: {e}")
+
+            self.track_parameter_count[name] += 1
+
+    def save_as_dataset(self, data_as_dict, savename, is_in) -> None:
+        """Save dictionnary of data to NetCDF
+
+        Note: Unknown types in the dictionnary won't be saved.
+        """
+        if savename not in self.stencils_or_class:
+            return
+
+        data_arrays = {}
+        for name, data in data_as_dict.items():
+            if dataclasses.is_dataclass(data):
+                for field in dataclasses.fields(data):
+                    data_arrays[f"{name}.{field.name}"] = self._to_xarray(
+                        getattr(data, field.name), field.name
+                    )
+            else:
+                data_arrays[name] = self._to_xarray(data, name)
+
+        call_count = (
+            self.calls_count[savename] if savename in self.calls_count.keys() else 0
+        )
+        path = pathlib.Path(f"{self.dir_name}/debug/savepoints/R{self.rank}/")
+        os.makedirs(path, exist_ok=True)
+        path = pathlib.Path(
+            f"{path}/{savename}-Call{call_count}-{'In' if is_in else 'Out'}.nc4"
+        )
+        try:
+            xr.Dataset(data_arrays).to_netcdf(path)
+        except ValueError as e:
+            ndsl_log.error(f"[DebugInfo] Failure to save {savename}: {e}")
+
+    def increment_call_count(self, savename: str):
+        """Increment the call count for this savename"""
+        if savename not in self.calls_count.keys():
+            self.calls_count[savename] = 0
+        self.calls_count[savename] += 1

ndsl/debug/tooling.py

@@ -0,0 +1,44 @@
+import inspect
+from functools import wraps
+from typing import Any, Callable
+
+from ndsl.debug.config import ndsl_debugger
+
+
+def instrument(func) -> Callable:
+    @wraps(func)
+    def wrapper(self, *args: Any, **kwargs: Any):
+        if ndsl_debugger is None:
+            return func(self, *args, **kwargs)
+        savename = func.__qualname__
+        params = inspect.signature(func).parameters
+        data_as_dict = {}
+
+        # Positional
+        positional_count = 0
+        for name, param in params.items():
+            if param.kind in (
+                inspect.Parameter.POSITIONAL_ONLY,
+                inspect.Parameter.POSITIONAL_OR_KEYWORD,
+            ):
+                if positional_count == 0:  # self
+                    positional_count += 1
+                    continue
+                if positional_count < len(args) + 1:
+                    data_as_dict[name] = args[positional_count - 1]
+                    positional_count += 1
+        # Keyword arguments
+        for name, value in kwargs.items():
+            if name in params:
+                data_as_dict[name] = value
+        if ndsl_debugger is not None:
+            ndsl_debugger.save_as_dataset(data_as_dict, func.__qualname__, is_in=True)
+            ndsl_debugger.track_data(data_as_dict, func.__qualname__, is_in=True)
+        r = func(self, *args, **kwargs)
+        if ndsl_debugger is not None:
+            ndsl_debugger.save_as_dataset(data_as_dict, func.__qualname__, is_in=False)
+            ndsl_debugger.track_data(data_as_dict, func.__qualname__, is_in=False)
+            ndsl_debugger.increment_call_count(savename)
+        return r
+
+    return wrapper

ndsl/dsl/stencil.py

@@ -21,12 +21,14 @@
 import numpy as np
 from gt4py.cartesian import gtscript
 from gt4py.cartesian.gtc.passes.oir_pipeline import DefaultPipeline, OirPipeline
+from gt4py.cartesian.stencil_object import StencilObject
 
 from ndsl.comm.comm_abc import Comm
 from ndsl.comm.communicator import Communicator
 from ndsl.comm.decomposition import block_waiting_for_compilation, unblock_waiting_tiles
 from ndsl.comm.mpi import MPI
 from ndsl.constants import X_DIM, X_DIMS, Y_DIM, Y_DIMS, Z_DIM, Z_DIMS
+from ndsl.debug import ndsl_debugger
 from ndsl.dsl.dace.orchestration import SDFGConvertible
 from ndsl.dsl.stencil_config import CompilationConfig, RunMode, StencilConfig
 from ndsl.dsl.typing import Float, Index3D, cast_to_index3d
@@ -295,10 +297,11 @@ def __init__(
             externals = {}
         self.externals = externals
         self._func_name = func.__name__
+        self._func_qualname = func.__qualname__
         stencil_kwargs = self.stencil_config.stencil_kwargs(
             skip_passes=skip_passes, func=func
         )
-        self.stencil_object = None
+        self.stencil_object: StencilObject | None = None
 
         self._argument_names = tuple(inspect.getfullargspec(func).args)
 
@@ -350,7 +353,7 @@ def __init__(
                 dtypes={float: Float},
                 **stencil_kwargs,
                 build_info=(build_info := {}),
-            )
+            )  # type: ignore
 
             if (
                 compilation_config.use_minimal_caching
@@ -384,20 +387,32 @@ def nothing_function(*args, **kwargs):
             setattr(self, "__call__", nothing_function)
 
     def __call__(self, *args, **kwargs) -> None:
+        # Verbose stencil execution
         if self.stencil_config.verbose:
             ndsl_log.debug(f"Running {self._func_name}")
+
+        # Marshal arguments
         args_list = list(args)
         _convert_quantities_to_storage(args_list, kwargs)
         args = tuple(args_list)
-
         args_as_kwargs = dict(zip(self._argument_names, args))
+
+        # Ranks comparison tool
         if self.comm is not None:
             differences = compare_ranks(self.comm, {**args_as_kwargs, **kwargs})
             if len(differences) > 0:
                 raise ValueError(
                     f"rank {self.comm.Get_rank()} has differences {differences} "
                     f"before calling {self._func_name}"
                 )
+
+        # Debugger actions if turned on
+        if ndsl_debugger:
+            all_args = args_as_kwargs | kwargs
+            ndsl_debugger.save_as_dataset(all_args, self._func_qualname, is_in=True)
+            ndsl_debugger.track_data(all_args, self._func_qualname, is_in=True)
+
+        # Execute stencil
         if self.stencil_config.compilation_config.validate_args:
             if __debug__ and "origin" in kwargs:
                 raise TypeError("origin cannot be passed to FrozenStencil call")
@@ -410,14 +425,23 @@ def __call__(self, *args, **kwargs) -> None:
                 domain=self.domain,
                 validate_args=True,
                 exec_info=self._timing_collector.exec_info,
-            )
+            )  # type: ignore
         else:
             self.stencil_object.run(
                 **args_as_kwargs,
                 **kwargs,
                 **self._stencil_run_kwargs,
                 exec_info=self._timing_collector.exec_info,
             )
+
+        # Debugger actions if turned on
+        if ndsl_debugger:
+            all_args = args_as_kwargs | kwargs
+            ndsl_debugger.save_as_dataset(all_args, self._func_qualname, is_in=False)
+            ndsl_debugger.track_data(all_args, self._func_qualname, is_in=False)
+            ndsl_debugger.increment_call_count(self._func_qualname)
+
+        # Ranks comparison tool
         if self.comm is not None:
             differences = compare_ranks(self.comm, {**args_as_kwargs, **kwargs})
             if len(differences) > 0:

ndsl/viz/__init__.py

@@ -0,0 +1,4 @@
+from .cube_sphere import plot_cube_sphere
+
+
+__all__ = ["plot_cube_sphere"]

ndsl/viz/cube_sphere.py

@@ -0,0 +1,36 @@
+import numpy as np
+from cartopy import crs as ccrs
+from matplotlib import pyplot as plt
+
+from ndsl import Quantity, ndsl_log
+from ndsl.comm.communicator import Communicator
+from ndsl.grid import GridData
+from ndsl.viz.fv3 import pcolormesh_cube
+
+
+def plot_cube_sphere(
+    quantity: Quantity,
+    k_level: int,
+    comm: Communicator,
+    grid_data: GridData,
+    save_to_path: str,
+):
+    if len(quantity.shape) < 2 or len(quantity.shape) > 3:
+        ndsl_log.error(
+            f"[Plot Cube] Can't plot quantity with shape == {quantity.shape}"
+        )
+        return
+
+    data = comm.gather(quantity)
+    lat = comm.gather(grid_data.lat)
+    lon = comm.gather(grid_data.lon)
+
+    if comm.rank == 0:
+        fig, ax = plt.subplots(1, 1, subplot_kw={"projection": ccrs.Robinson()})
+        pcolormesh_cube(
+            lat.view[:] * 180.0 / np.pi,
+            lon.view[:] * 180.0 / np.pi,
+            data.view[:] if len(data.shape) == 3 else data.view[:, :, :, k_level],
+            ax=ax,
+        )
+        fig.savefig(save_to_path)

ndsl/viz/fv3/README.md

@@ -0,0 +1,14 @@
+# Acknowledgment
+
+This code was lifted from <https://github.com/ai2cm/fv3net> and developped by AI2 under the MIT license (see below).
+
+## MIT License
+
+The MIT License (MIT)
+Copyright (c) 2019, The Allen Institute for Artificial Intelligence
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.