feat(tidy3d): FXC-3343-all-port-excitation-with-post-normalization-parallel-adjoint

CHANGELOG.md

@@ -19,6 +19,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - 1D lumped elements (with zero lateral extent) are no longer allowed. Use a small finite lateral extent (e.g., `1e-6`) instead.
 - Added `GaussianOverlapMonitor` and `AstigmaticGaussianOverlapMonitor` for decomposing electromagnetic fields onto Gaussian beam profiles.
 - Added `GaussianPort` and `AstigmaticGaussianPort` for S-matrix calculations using Gaussian beam sources and overlap monitors.
+- Added parallel adjoint scheduling (concurrent with forward simulation) for suitable monitors via `config.adjoint.parallel_all_port` (if `local gradient=True`).
 
 ### Changed
 - `ModeSortSpec.sort_key` is now required with a default of `"n_eff"` (previously optional with `None` default). `ModeSortSpec.sort_order` is now optional with a default of `None`, which automatically selects the natural order based on `sort_key` and `sort_reference`: ascending when a reference is provided (closest first), otherwise descending for `n_eff` and polarization fractions (higher values first), ascending for `k_eff` and `mode_area` (lower values first).

docs/configuration/reference.rst

@@ -125,6 +125,10 @@ remote defaults and emits a warning reminding you to enable local gradients.
      - ``"adjoint_data"``
      - Yes
      - Directory (relative to the working directory) where intermediate gradient artifacts are stored when ``local_gradient`` is enabled.
+   * - ``parallel_all_port``
+     - ``False``
+     - Yes
+     - Launch canonical adjoint simulations for supported monitors in parallel with the forward solve when local gradients are enabled.
    * - ``gradient_precision``
      - ``"single"``
      - No

tests/test_components/autograd/test_autograd.py

@@ -282,9 +282,16 @@ def emulated_run_async_fwd(simulations, **run_kwargs) -> td.SimulationData:
             batch_data_orig, task_ids_fwd = {}, {}
             sim_fields_keys_dict = run_kwargs.pop("sim_fields_keys_dict", None)
             for task_name, simulation in simulations.items():
+                run_kwargs_task = dict(run_kwargs)
                 if sim_fields_keys_dict is not None:
-                    run_kwargs["sim_fields_keys"] = sim_fields_keys_dict[task_name]
-                sim_data_orig, task_name_fwd = emulated_run_fwd(simulation, task_name, **run_kwargs)
+                    sim_fields_keys = sim_fields_keys_dict.get(task_name)
+                    if sim_fields_keys is not None and "_parallel_adj_" not in task_name:
+                        run_kwargs_task["sim_fields_keys"] = sim_fields_keys
+                    else:
+                        run_kwargs_task.pop("sim_fields_keys", None)
+                sim_data_orig, task_name_fwd = emulated_run_fwd(
+                    simulation, task_name, **run_kwargs_task
+                )
                 batch_data_orig[task_name] = sim_data_orig
                 task_ids_fwd[task_name] = task_name_fwd
 
@@ -3005,7 +3012,7 @@ def objective(args):
 
     # the following will raise a warning (and fail) if the dispersive material
     # model is called without a frequency
-    with AssertLogLevel("INFO"):
+    with AssertLogLevel("WARNING"):
         grad = ag.grad(objective)(params0)