From b69c8c5e7ed030259cef6ae2f00d2f577b6134e5 Mon Sep 17 00:00:00 2001
From: Max <max.lindqvist@ipp.mpg.de>
Date: Wed, 7 Oct 2026 22:55:46 +0200
Subject: [PATCH] Declare kernel outputs for PyccelKernels

Each kernel folder's __init__.py declares the arguments its kernel writes to
(OUTPUTS, passed to cunumpy as host_options={"outputs": OUTPUTS}); the
PyccelKernel wraps in feec/mass.py, boundary_mass.py and
basis_projection_ops.py pass outputs= too. The parity tests compare the
declared outputs; new tests check the declarations on the NumPy backend
and in the CPU emulation of the CUDA kernels.

Solves #342.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 CUDA_STRATEGY.md                              | 10 +++++
 .../eval_spline_mpi_markers/__init__.py       |  5 ++-
 .../eval_spline_mpi_matrix/__init__.py        |  5 ++-
 .../__init__.py                               |  5 ++-
 src/struphy/feec/basis_projection_ops.py      |  1 +
 src/struphy/feec/boundary_mass.py             |  4 +-
 .../feec/kernels/hybrid_weight/__init__.py    |  5 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../kernels/get_dofs_local_3_form/__init__.py |  5 ++-
 .../__init__.py                               |  5 ++-
 .../kernels/solve_local_main_loop/__init__.py |  5 ++-
 .../__init__.py                               |  5 ++-
 src/struphy/feec/mass.py                      |  5 ++-
 .../kernels/kernel_evaluate/__init__.py       |  5 ++-
 .../kernels/kernel_evaluate_pic/__init__.py   |  5 ++-
 .../kernels/kernel_pullpush/__init__.py       |  5 ++-
 .../kernels/kernel_pullpush_pic/__init__.py   |  5 ++-
 .../kernels/cc_lin_mhd_5d_D/__init__.py       |  5 ++-
 .../kernels/cc_lin_mhd_5d_M/__init__.py       |  5 ++-
 .../kernels/cc_lin_mhd_5d_curlb/__init__.py   |  5 ++-
 .../kernels/cc_lin_mhd_5d_gradB/__init__.py   |  5 ++-
 .../cc_lin_mhd_5d_gradB_dg/__init__.py        |  5 ++-
 .../cc_lin_mhd_5d_gradB_dg_init/__init__.py   |  5 ++-
 .../kernels/cc_lin_mhd_6d_1/__init__.py       |  5 ++-
 .../kernels/cc_lin_mhd_6d_2/__init__.py       |  5 ++-
 .../kernels/charge_density_0form/__init__.py  |  5 ++-
 .../kernels/div_u_weak_1form/__init__.py      |  5 ++-
 .../kernels/gc_density_0form/__init__.py      |  5 ++-
 .../kernels/gc_mag_density_0form/__init__.py  |  5 ++-
 .../kernels/linear_vlasov_ampere/__init__.py  |  5 ++-
 .../kernels/pc_lin_mhd_6d/__init__.py         |  7 ++-
 .../kernels/pc_lin_mhd_6d_full/__init__.py    |  5 ++-
 .../kernels/vlasov_maxwell/__init__.py        |  5 ++-
 .../canonical_kinetic_particles/__init__.py   |  5 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../kernels/eval_energy_5d/__init__.py        |  5 ++-
 .../kernels/eval_gradB_ediff/__init__.py      |  5 ++-
 .../eval_guiding_center_from_6d/__init__.py   |  5 ++-
 .../__init__.py                               |  5 ++-
 .../eval_magnetic_energy_PBb/__init__.py      |  5 ++-
 .../eval_magnetic_moment_5d/__init__.py       |  5 ++-
 .../kernels/thermal_energy/__init__.py        |  5 ++-
 .../pushing/kernels/bstar_2form/__init__.py   |  5 ++-
 .../kernels/bstar_parallel_3form/__init__.py  |  5 ++-
 .../driftkinetic_hamiltonian/__init__.py      |  5 ++-
 .../grad_driftkinetic_hamiltonian/__init__.py |  5 ++-
 .../kernels/push_bxu_H1vec/__init__.py        |  5 ++-
 .../kernels/push_bxu_Hcurl/__init__.py        |  5 ++-
 .../pushing/kernels/push_bxu_Hdiv/__init__.py |  5 ++-
 .../kernels/push_bxu_Hdiv_pauli/__init__.py   |  5 ++-
 .../__init__.py                               |  7 ++-
 .../kernels/push_eta_stage/__init__.py        |  5 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../__init__.py                               |  7 ++-
 .../kernels/push_gc_cc_J1_H1vec/__init__.py   |  5 ++-
 .../kernels/push_gc_cc_J1_Hcurl/__init__.py   |  5 ++-
 .../kernels/push_gc_cc_J1_Hdiv/__init__.py    |  5 ++-
 .../kernels/push_gc_cc_J2_dg_Hdiv/__init__.py |  5 ++-
 .../push_gc_cc_J2_dg_init_Hdiv/__init__.py    |  5 ++-
 .../push_gc_cc_J2_stage_H1vec/__init__.py     |  5 ++-
 .../push_gc_cc_J2_stage_Hdiv/__init__.py      |  5 ++-
 .../pushing/kernels/push_pc_GXu/__init__.py   |  5 ++-
 .../kernels/push_pc_GXu_full/__init__.py      |  5 ++-
 .../push_pc_eta_stage_H1vec/__init__.py       |  5 ++-
 .../push_pc_eta_stage_Hcurl/__init__.py       |  5 ++-
 .../push_pc_eta_stage_Hdiv/__init__.py        |  5 ++-
 .../push_random_diffusion_stage/__init__.py   |  5 ++-
 .../kernels/push_v_sph_pressure/__init__.py   |  5 ++-
 .../push_v_sph_pressure_ideal_gas/__init__.py |  5 ++-
 .../kernels/push_v_viscosity/__init__.py      |  5 ++-
 .../kernels/push_v_with_efield/__init__.py    |  5 ++-
 .../kernels/push_vxb_analytic/__init__.py     |  5 ++-
 .../kernels/push_vxb_implicit/__init__.py     |  5 ++-
 .../__init__.py                               |  5 ++-
 .../pic/pushing/kernels/reflect/__init__.py   |  5 ++-
 .../kernels/sph_isotherm_kappa/__init__.py    |  5 ++-
 .../sph_mean_velocity_coeffs/__init__.py      |  5 ++-
 .../kernels/sph_pressure_coeffs/__init__.py   |  5 ++-
 .../kernels/sph_viscosity_tensor/__init__.py  |  5 ++-
 .../pushing/kernels/unit_b_1form/__init__.py  |  5 ++-
 .../box_based_evaluation_flat/__init__.py     |  5 ++-
 .../box_based_evaluation_meshgrid/__init__.py |  5 ++-
 .../kernels/naive_evaluation_flat/__init__.py |  5 ++-
 .../naive_evaluation_meshgrid/__init__.py     |  5 ++-
 src/struphy/pic/tests/cuda_parity_cases.py    | 33 +++++++++++++-
 src/struphy/pic/tests/test_cuda_emulation.py  | 38 +++++-----------
 src/struphy/pic/tests/test_cuda_parity.py     | 44 +++++++++++++++++--
 97 files changed, 493 insertions(+), 124 deletions(-)

diff --git a/CUDA_STRATEGY.md b/CUDA_STRATEGY.md
index daf0d25f0..55498ec72 100644
--- a/CUDA_STRATEGY.md
+++ b/CUDA_STRATEGY.md
@@ -537,6 +537,16 @@ They are replaced by one test module, `pic/tests/cuda_parity_cases.py`: `PARITY_
 `assert_kernels_agree` (one test per case), `test_cuda_emulation.py` to the CPU emulation, and
 `test_cuda_kernels_have_parity_cases` checks that the CUDA kernels and the parity cases match.
 
+**Declared kernel outputs (#342).** Each kernel folder's `__init__.py` declares the arguments its kernel writes
+to, by position, as `OUTPUTS` and passes them to cunumpy (`Kernel.from_folder(..., host_options={"outputs":
+OUTPUTS})`); the `PyccelKernel(...)` wraps in `feec/mass.py`, `feec/boundary_mass.py` and
+`feec/basis_projection_ops.py` pass `outputs=` too. Scratch arrays inside argument objects (`DerhamArguments.bn1`,
+...) are not outputs. On the CuPy fallback only the outputs are copied back to the device; the GPU parity test
+(`assert_kernels_agree`) compares only the outputs; `test_declared_outputs` checks on the NumPy backend that a kernel
+changes its outputs and no other argument, and the CPU emulation test checks the same for the CUDA kernel.
+`KernelCatalog.from_package` does not see the folders' `OUTPUTS`, so `test_cuda_parity` takes the kernels from the
+folders (`DECLARED`). A new kernel folder declares `OUTPUTS` (`test_kernels_declare_outputs`).
+
 ## PR 18 implementation notes
 
 Geometry evaluation runs on the GPU for every analytic mapping (`kind_map` 10–12, 20–22, 30–32).
diff --git a/src/struphy/bsplines/kernels/eval_spline_mpi_markers/__init__.py b/src/struphy/bsplines/kernels/eval_spline_mpi_markers/__init__.py
index faedc1370..51a9b6d32 100644
--- a/src/struphy/bsplines/kernels/eval_spline_mpi_markers/__init__.py
+++ b/src/struphy/bsplines/kernels/eval_spline_mpi_markers/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_spline_mpi_markers = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: values
+OUTPUTS = (8,)
+
+eval_spline_mpi_markers = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/bsplines/kernels/eval_spline_mpi_matrix/__init__.py b/src/struphy/bsplines/kernels/eval_spline_mpi_matrix/__init__.py
index 4b453f794..b501e532c 100644
--- a/src/struphy/bsplines/kernels/eval_spline_mpi_matrix/__init__.py
+++ b/src/struphy/bsplines/kernels/eval_spline_mpi_matrix/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_spline_mpi_matrix = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: values
+OUTPUTS = (10,)
+
+eval_spline_mpi_matrix = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/bsplines/kernels/eval_spline_mpi_sparse_meshgrid/__init__.py b/src/struphy/bsplines/kernels/eval_spline_mpi_sparse_meshgrid/__init__.py
index f7ae87826..8600c77fa 100644
--- a/src/struphy/bsplines/kernels/eval_spline_mpi_sparse_meshgrid/__init__.py
+++ b/src/struphy/bsplines/kernels/eval_spline_mpi_sparse_meshgrid/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_spline_mpi_sparse_meshgrid = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: values
+OUTPUTS = (10,)
+
+eval_spline_mpi_sparse_meshgrid = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/feec/basis_projection_ops.py b/src/struphy/feec/basis_projection_ops.py
index ba2c29dec..d87782299 100644
--- a/src/struphy/feec/basis_projection_ops.py
+++ b/src/struphy/feec/basis_projection_ops.py
@@ -2084,6 +2084,7 @@ def assemble(self, weights=None):
                             basis_projection_kernels,
                             "assemble_dofs_for_weighted_basisfuns_" + str(V.ldim) + "d",
                         ),
+                        outputs=(0,),  # writes the matrix data
                     )
 
                     logger.debug(f"Assemble block {i, j}")
diff --git a/src/struphy/feec/boundary_mass.py b/src/struphy/feec/boundary_mass.py
index 9c2953245..1382cd98d 100644
--- a/src/struphy/feec/boundary_mass.py
+++ b/src/struphy/feec/boundary_mass.py
@@ -173,8 +173,8 @@ def __init__(
         # for each active face, precompute per-space surface quadrature/geometric data
         self._setup_surface_data()
 
-        # load assembly kernel
-        self._assembly_kernel = PyccelKernel(mass_kernels.surface_kernel_3d_mat)
+        # load assembly kernel (writes the matrix data, its last argument)
+        self._assembly_kernel = PyccelKernel(mass_kernels.surface_kernel_3d_mat, outputs=(-1,))
 
         self.assemble()
 
diff --git a/src/struphy/feec/kernels/hybrid_weight/__init__.py b/src/struphy/feec/kernels/hybrid_weight/__init__.py
index 7af87b677..c190ed2f5 100644
--- a/src/struphy/feec/kernels/hybrid_weight/__init__.py
+++ b/src/struphy/feec/kernels/hybrid_weight/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-hybrid_weight = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: data1, data2, data3
+OUTPUTS = (15, 16, 17)
+
+hybrid_weight = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/feec/local_projectors/kernels/get_dofs_local_1_form_ec_component/__init__.py b/src/struphy/feec/local_projectors/kernels/get_dofs_local_1_form_ec_component/__init__.py
index 7ebfc1574..039630c75 100644
--- a/src/struphy/feec/local_projectors/kernels/get_dofs_local_1_form_ec_component/__init__.py
+++ b/src/struphy/feec/local_projectors/kernels/get_dofs_local_1_form_ec_component/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-get_dofs_local_1_form_ec_component = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: f_eval_aux
+OUTPUTS = (2,)
+
+get_dofs_local_1_form_ec_component = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/feec/local_projectors/kernels/get_dofs_local_1_form_ec_component_weighted/__init__.py b/src/struphy/feec/local_projectors/kernels/get_dofs_local_1_form_ec_component_weighted/__init__.py
index 9002340c1..a3d0d785c 100644
--- a/src/struphy/feec/local_projectors/kernels/get_dofs_local_1_form_ec_component_weighted/__init__.py
+++ b/src/struphy/feec/local_projectors/kernels/get_dofs_local_1_form_ec_component_weighted/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-get_dofs_local_1_form_ec_component_weighted = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: f_eval_aux
+OUTPUTS = (6,)
+
+get_dofs_local_1_form_ec_component_weighted = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/feec/local_projectors/kernels/get_dofs_local_2_form_ec_component/__init__.py b/src/struphy/feec/local_projectors/kernels/get_dofs_local_2_form_ec_component/__init__.py
index 04d7262b0..1a3e78ed2 100644
--- a/src/struphy/feec/local_projectors/kernels/get_dofs_local_2_form_ec_component/__init__.py
+++ b/src/struphy/feec/local_projectors/kernels/get_dofs_local_2_form_ec_component/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-get_dofs_local_2_form_ec_component = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: f_eval_aux
+OUTPUTS = (2,)
+
+get_dofs_local_2_form_ec_component = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/feec/local_projectors/kernels/get_dofs_local_2_form_ec_component_weighted/__init__.py b/src/struphy/feec/local_projectors/kernels/get_dofs_local_2_form_ec_component_weighted/__init__.py
index 2c27eab86..091907165 100644
--- a/src/struphy/feec/local_projectors/kernels/get_dofs_local_2_form_ec_component_weighted/__init__.py
+++ b/src/struphy/feec/local_projectors/kernels/get_dofs_local_2_form_ec_component_weighted/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-get_dofs_local_2_form_ec_component_weighted = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: f_eval_aux
+OUTPUTS = (7,)
+
+get_dofs_local_2_form_ec_component_weighted = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/feec/local_projectors/kernels/get_dofs_local_3_form/__init__.py b/src/struphy/feec/local_projectors/kernels/get_dofs_local_3_form/__init__.py
index e684886ec..692975a95 100644
--- a/src/struphy/feec/local_projectors/kernels/get_dofs_local_3_form/__init__.py
+++ b/src/struphy/feec/local_projectors/kernels/get_dofs_local_3_form/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-get_dofs_local_3_form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: f_eval
+OUTPUTS = (2,)
+
+get_dofs_local_3_form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/feec/local_projectors/kernels/get_dofs_local_3_form_weighted/__init__.py b/src/struphy/feec/local_projectors/kernels/get_dofs_local_3_form_weighted/__init__.py
index 1c6556895..137bb5a33 100644
--- a/src/struphy/feec/local_projectors/kernels/get_dofs_local_3_form_weighted/__init__.py
+++ b/src/struphy/feec/local_projectors/kernels/get_dofs_local_3_form_weighted/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-get_dofs_local_3_form_weighted = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: f_eval
+OUTPUTS = (8,)
+
+get_dofs_local_3_form_weighted = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/feec/local_projectors/kernels/solve_local_main_loop/__init__.py b/src/struphy/feec/local_projectors/kernels/solve_local_main_loop/__init__.py
index b973111c4..0ea091d02 100644
--- a/src/struphy/feec/local_projectors/kernels/solve_local_main_loop/__init__.py
+++ b/src/struphy/feec/local_projectors/kernels/solve_local_main_loop/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-solve_local_main_loop = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: out
+OUTPUTS = (2,)
+
+solve_local_main_loop = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/feec/local_projectors/kernels/solve_local_main_loop_weighted/__init__.py b/src/struphy/feec/local_projectors/kernels/solve_local_main_loop_weighted/__init__.py
index 9fd6131fe..5f2b7aea7 100644
--- a/src/struphy/feec/local_projectors/kernels/solve_local_main_loop_weighted/__init__.py
+++ b/src/struphy/feec/local_projectors/kernels/solve_local_main_loop_weighted/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-solve_local_main_loop_weighted = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: out
+OUTPUTS = (8,)
+
+solve_local_main_loop_weighted = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/feec/mass.py b/src/struphy/feec/mass.py
index e56a19af0..0e229dd32 100644
--- a/src/struphy/feec/mass.py
+++ b/src/struphy/feec/mass.py
@@ -1935,6 +1935,7 @@ def _load_assembly_kernel(self):
                 mass_kernels,
                 "kernel_" + str(self._V.ldim) + "d_mat",
             ),
+            outputs=(-1,),  # writes the matrix data
         )
 
     @property
@@ -2570,7 +2571,7 @@ def eval_quad(self, W, coeffs, out=None):
                 assert isinstance(out, (list, tuple))
 
         # load assembly kernel
-        kernel = PyccelKernel(getattr(mass_kernels, "kernel_" + str(W.ldim) + "d_eval"))
+        kernel = PyccelKernel(getattr(mass_kernels, "kernel_" + str(W.ldim) + "d_eval"), outputs=(-1,))  # writes `out`
 
         # loop over components
         for a, wspace in enumerate(Wspaces):
@@ -2673,6 +2674,7 @@ def __init__(self, derham, V, W, weights=None, nquads=None):
                 mass_kernels,
                 "kernel_" + str(self._V.ldim) + "d_matrixfree",
             ),
+            outputs=(-2,),  # writes the data of `out` (the last argument is the input vector)
         )
 
         self._diag_kernel = PyccelKernel(
@@ -2680,6 +2682,7 @@ def __init__(self, derham, V, W, weights=None, nquads=None):
                 mass_kernels,
                 "kernel_" + str(self._V.ldim) + "d_diag",
             ),
+            outputs=(-1,),  # writes the diagonal data
         )
 
         # temporary with ghost regions for the diagonal (contributions to other processes are exchanged)
diff --git a/src/struphy/geometry/kernels/kernel_evaluate/__init__.py b/src/struphy/geometry/kernels/kernel_evaluate/__init__.py
index 2dce667df..86641f545 100644
--- a/src/struphy/geometry/kernels/kernel_evaluate/__init__.py
+++ b/src/struphy/geometry/kernels/kernel_evaluate/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-kernel_evaluate = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat_f
+OUTPUTS = (5,)
+
+kernel_evaluate = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/geometry/kernels/kernel_evaluate_pic/__init__.py b/src/struphy/geometry/kernels/kernel_evaluate_pic/__init__.py
index 84ab2776b..232efa05b 100644
--- a/src/struphy/geometry/kernels/kernel_evaluate_pic/__init__.py
+++ b/src/struphy/geometry/kernels/kernel_evaluate_pic/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-kernel_evaluate_pic = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat_f
+OUTPUTS = (3,)
+
+kernel_evaluate_pic = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/geometry/kernels/kernel_pullpush/__init__.py b/src/struphy/geometry/kernels/kernel_pullpush/__init__.py
index 3d3ca4de1..6dfb2c205 100644
--- a/src/struphy/geometry/kernels/kernel_pullpush/__init__.py
+++ b/src/struphy/geometry/kernels/kernel_pullpush/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-kernel_pullpush = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: out
+OUTPUTS = (8,)
+
+kernel_pullpush = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/geometry/kernels/kernel_pullpush_pic/__init__.py b/src/struphy/geometry/kernels/kernel_pullpush_pic/__init__.py
index 682cdf221..a23c932a7 100644
--- a/src/struphy/geometry/kernels/kernel_pullpush_pic/__init__.py
+++ b/src/struphy/geometry/kernels/kernel_pullpush_pic/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-kernel_pullpush_pic = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: out
+OUTPUTS = (5,)
+
+kernel_pullpush_pic = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_D/__init__.py b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_D/__init__.py
index 27af54854..fd8ecbf4d 100644
--- a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_D/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_D/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-cc_lin_mhd_5d_D = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat12, mat13, mat23
+OUTPUTS = (3, 4, 5)
+
+cc_lin_mhd_5d_D = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_M/__init__.py b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_M/__init__.py
index b2f152444..65309988a 100644
--- a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_M/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_M/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-cc_lin_mhd_5d_M = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: vec1, vec2, vec3
+OUTPUTS = (9, 10, 11)
+
+cc_lin_mhd_5d_M = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_curlb/__init__.py b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_curlb/__init__.py
index c8bfc2422..7b158ed13 100644
--- a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_curlb/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_curlb/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-cc_lin_mhd_5d_curlb = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat11 to vec3
+OUTPUTS = tuple(range(3, 12))
+
+cc_lin_mhd_5d_curlb = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB/__init__.py b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB/__init__.py
index 33cdbb970..d7519aae2 100644
--- a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-cc_lin_mhd_5d_gradB = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: vec1, vec2, vec3
+OUTPUTS = (9, 10, 11)
+
+cc_lin_mhd_5d_gradB = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB_dg/__init__.py b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB_dg/__init__.py
index 15ead0123..7fe56051a 100644
--- a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB_dg/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB_dg/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-cc_lin_mhd_5d_gradB_dg = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: vec1, vec2, vec3
+OUTPUTS = (3, 4, 5)
+
+cc_lin_mhd_5d_gradB_dg = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB_dg_init/__init__.py b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB_dg_init/__init__.py
index 464b9a7a9..61d48afd2 100644
--- a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB_dg_init/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_5d_gradB_dg_init/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-cc_lin_mhd_5d_gradB_dg_init = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: vec1, vec2, vec3
+OUTPUTS = (3, 4, 5)
+
+cc_lin_mhd_5d_gradB_dg_init = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_6d_1/__init__.py b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_6d_1/__init__.py
index be58bb73f..2deeae23b 100644
--- a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_6d_1/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_6d_1/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-cc_lin_mhd_6d_1 = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat12, mat13, mat23
+OUTPUTS = (3, 4, 5)
+
+cc_lin_mhd_6d_1 = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_6d_2/__init__.py b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_6d_2/__init__.py
index 7aaac64c9..daad46474 100644
--- a/src/struphy/pic/accumulation/kernels/cc_lin_mhd_6d_2/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/cc_lin_mhd_6d_2/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-cc_lin_mhd_6d_2 = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat11 to vec3
+OUTPUTS = tuple(range(3, 12))
+
+cc_lin_mhd_6d_2 = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/charge_density_0form/__init__.py b/src/struphy/pic/accumulation/kernels/charge_density_0form/__init__.py
index b136492c0..aae4e3397 100644
--- a/src/struphy/pic/accumulation/kernels/charge_density_0form/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/charge_density_0form/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-charge_density_0form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: vec
+OUTPUTS = (3,)
+
+charge_density_0form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/div_u_weak_1form/__init__.py b/src/struphy/pic/accumulation/kernels/div_u_weak_1form/__init__.py
index a53a9782b..9898dee75 100644
--- a/src/struphy/pic/accumulation/kernels/div_u_weak_1form/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/div_u_weak_1form/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-div_u_weak_1form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: vec1, vec2, vec3
+OUTPUTS = (3, 4, 5)
+
+div_u_weak_1form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/gc_density_0form/__init__.py b/src/struphy/pic/accumulation/kernels/gc_density_0form/__init__.py
index 9ba544769..3fbe54e81 100644
--- a/src/struphy/pic/accumulation/kernels/gc_density_0form/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/gc_density_0form/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-gc_density_0form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: vec
+OUTPUTS = (3,)
+
+gc_density_0form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/gc_mag_density_0form/__init__.py b/src/struphy/pic/accumulation/kernels/gc_mag_density_0form/__init__.py
index d7bfa1346..f207c7756 100644
--- a/src/struphy/pic/accumulation/kernels/gc_mag_density_0form/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/gc_mag_density_0form/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-gc_mag_density_0form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: vec
+OUTPUTS = (3,)
+
+gc_mag_density_0form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/linear_vlasov_ampere/__init__.py b/src/struphy/pic/accumulation/kernels/linear_vlasov_ampere/__init__.py
index 0fed66c87..035b53986 100644
--- a/src/struphy/pic/accumulation/kernels/linear_vlasov_ampere/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/linear_vlasov_ampere/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-linear_vlasov_ampere = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat11 to vec3
+OUTPUTS = tuple(range(3, 12))
+
+linear_vlasov_ampere = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/pc_lin_mhd_6d/__init__.py b/src/struphy/pic/accumulation/kernels/pc_lin_mhd_6d/__init__.py
index ddb2e6fb4..616275df4 100644
--- a/src/struphy/pic/accumulation/kernels/pc_lin_mhd_6d/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/pc_lin_mhd_6d/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-pc_lin_mhd_6d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat11_11, mat12_11, mat13_11, mat22_11, mat23_11, mat33_11,
+# mat11_12, mat12_12, mat13_12, mat22_12, mat23_12, mat33_12, mat11_22, mat12_22, mat13_22, mat22_22, mat23_22,
+# mat33_22, vec1_1, vec2_1, vec3_1, vec1_2, vec2_2, vec3_2
+OUTPUTS = (3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 21, 22, 23, 24, 25, 26, 39, 40, 41, 42, 43, 44)
+
+pc_lin_mhd_6d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/pc_lin_mhd_6d_full/__init__.py b/src/struphy/pic/accumulation/kernels/pc_lin_mhd_6d_full/__init__.py
index 97a011e8b..456e05263 100644
--- a/src/struphy/pic/accumulation/kernels/pc_lin_mhd_6d_full/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/pc_lin_mhd_6d_full/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-pc_lin_mhd_6d_full = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat11_11 to vec3_3
+OUTPUTS = tuple(range(3, 48))
+
+pc_lin_mhd_6d_full = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/accumulation/kernels/vlasov_maxwell/__init__.py b/src/struphy/pic/accumulation/kernels/vlasov_maxwell/__init__.py
index 8cfb5fc55..19cee8249 100644
--- a/src/struphy/pic/accumulation/kernels/vlasov_maxwell/__init__.py
+++ b/src/struphy/pic/accumulation/kernels/vlasov_maxwell/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-vlasov_maxwell = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: mat11 to vec3
+OUTPUTS = tuple(range(3, 12))
+
+vlasov_maxwell = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/diagnostics/kernels/canonical_kinetic_particles/__init__.py b/src/struphy/pic/diagnostics/kernels/canonical_kinetic_particles/__init__.py
index 48f837548..c7f49bbb2 100644
--- a/src/struphy/pic/diagnostics/kernels/canonical_kinetic_particles/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/canonical_kinetic_particles/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-canonical_kinetic_particles = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: res
+OUTPUTS = (0,)
+
+canonical_kinetic_particles = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/diagnostics/kernels/eval_canonical_toroidal_moment_5d/__init__.py b/src/struphy/pic/diagnostics/kernels/eval_canonical_toroidal_moment_5d/__init__.py
index 7c0d66d24..a826c4798 100644
--- a/src/struphy/pic/diagnostics/kernels/eval_canonical_toroidal_moment_5d/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/eval_canonical_toroidal_moment_5d/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_canonical_toroidal_moment_5d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: markers
+OUTPUTS = (0,)
+
+eval_canonical_toroidal_moment_5d = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/diagnostics/kernels/eval_canonical_toroidal_moment_6d/__init__.py b/src/struphy/pic/diagnostics/kernels/eval_canonical_toroidal_moment_6d/__init__.py
index fb815faa4..3bd33e1af 100644
--- a/src/struphy/pic/diagnostics/kernels/eval_canonical_toroidal_moment_6d/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/eval_canonical_toroidal_moment_6d/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_canonical_toroidal_moment_6d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: markers
+OUTPUTS = (0,)
+
+eval_canonical_toroidal_moment_6d = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/diagnostics/kernels/eval_energy_5d/__init__.py b/src/struphy/pic/diagnostics/kernels/eval_energy_5d/__init__.py
index 2e2c02830..da11b29d1 100644
--- a/src/struphy/pic/diagnostics/kernels/eval_energy_5d/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/eval_energy_5d/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_energy_5d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: markers
+OUTPUTS = (0,)
+
+eval_energy_5d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/diagnostics/kernels/eval_gradB_ediff/__init__.py b/src/struphy/pic/diagnostics/kernels/eval_gradB_ediff/__init__.py
index 80e6a8201..e35d2ae9f 100644
--- a/src/struphy/pic/diagnostics/kernels/eval_gradB_ediff/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/eval_gradB_ediff/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_gradB_ediff = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (0,)
+
+eval_gradB_ediff = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/diagnostics/kernels/eval_guiding_center_from_6d/__init__.py b/src/struphy/pic/diagnostics/kernels/eval_guiding_center_from_6d/__init__.py
index 161f4bf10..984ff87e6 100644
--- a/src/struphy/pic/diagnostics/kernels/eval_guiding_center_from_6d/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/eval_guiding_center_from_6d/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_guiding_center_from_6d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: markers
+OUTPUTS = (0,)
+
+eval_guiding_center_from_6d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/diagnostics/kernels/eval_magnetic_background_energy/__init__.py b/src/struphy/pic/diagnostics/kernels/eval_magnetic_background_energy/__init__.py
index 4402687bc..6f4ba75f0 100644
--- a/src/struphy/pic/diagnostics/kernels/eval_magnetic_background_energy/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/eval_magnetic_background_energy/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_magnetic_background_energy = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: markers
+OUTPUTS = (0,)
+
+eval_magnetic_background_energy = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/diagnostics/kernels/eval_magnetic_energy_PBb/__init__.py b/src/struphy/pic/diagnostics/kernels/eval_magnetic_energy_PBb/__init__.py
index b6475cd0c..90e1014be 100644
--- a/src/struphy/pic/diagnostics/kernels/eval_magnetic_energy_PBb/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/eval_magnetic_energy_PBb/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_magnetic_energy_PBb = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: markers
+OUTPUTS = (0,)
+
+eval_magnetic_energy_PBb = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/diagnostics/kernels/eval_magnetic_moment_5d/__init__.py b/src/struphy/pic/diagnostics/kernels/eval_magnetic_moment_5d/__init__.py
index 02ae3acd8..f88bd9265 100644
--- a/src/struphy/pic/diagnostics/kernels/eval_magnetic_moment_5d/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/eval_magnetic_moment_5d/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-eval_magnetic_moment_5d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: markers
+OUTPUTS = (0,)
+
+eval_magnetic_moment_5d = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/diagnostics/kernels/thermal_energy/__init__.py b/src/struphy/pic/diagnostics/kernels/thermal_energy/__init__.py
index 326b75882..3866e6cb3 100644
--- a/src/struphy/pic/diagnostics/kernels/thermal_energy/__init__.py
+++ b/src/struphy/pic/diagnostics/kernels/thermal_energy/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-thermal_energy = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: res
+OUTPUTS = (0,)
+
+thermal_energy = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/bstar_2form/__init__.py b/src/struphy/pic/pushing/kernels/bstar_2form/__init__.py
index cea1e83ad..a3b28a3b4 100644
--- a/src/struphy/pic/pushing/kernels/bstar_2form/__init__.py
+++ b/src/struphy/pic/pushing/kernels/bstar_2form/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-bstar_2form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+bstar_2form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/bstar_parallel_3form/__init__.py b/src/struphy/pic/pushing/kernels/bstar_parallel_3form/__init__.py
index 20ce21b71..1380958bd 100644
--- a/src/struphy/pic/pushing/kernels/bstar_parallel_3form/__init__.py
+++ b/src/struphy/pic/pushing/kernels/bstar_parallel_3form/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-bstar_parallel_3form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+bstar_parallel_3form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/driftkinetic_hamiltonian/__init__.py b/src/struphy/pic/pushing/kernels/driftkinetic_hamiltonian/__init__.py
index ae1cf069d..f297dc991 100644
--- a/src/struphy/pic/pushing/kernels/driftkinetic_hamiltonian/__init__.py
+++ b/src/struphy/pic/pushing/kernels/driftkinetic_hamiltonian/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-driftkinetic_hamiltonian = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+driftkinetic_hamiltonian = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/grad_driftkinetic_hamiltonian/__init__.py b/src/struphy/pic/pushing/kernels/grad_driftkinetic_hamiltonian/__init__.py
index 1ebe8b1a6..a71e77859 100644
--- a/src/struphy/pic/pushing/kernels/grad_driftkinetic_hamiltonian/__init__.py
+++ b/src/struphy/pic/pushing/kernels/grad_driftkinetic_hamiltonian/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-grad_driftkinetic_hamiltonian = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+grad_driftkinetic_hamiltonian = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_bxu_H1vec/__init__.py b/src/struphy/pic/pushing/kernels/push_bxu_H1vec/__init__.py
index a42e017cd..2bc1dc3d6 100644
--- a/src/struphy/pic/pushing/kernels/push_bxu_H1vec/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_bxu_H1vec/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_bxu_H1vec = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_bxu_H1vec = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_bxu_Hcurl/__init__.py b/src/struphy/pic/pushing/kernels/push_bxu_Hcurl/__init__.py
index 8f07d5295..952f71d8b 100644
--- a/src/struphy/pic/pushing/kernels/push_bxu_Hcurl/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_bxu_Hcurl/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_bxu_Hcurl = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_bxu_Hcurl = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_bxu_Hdiv/__init__.py b/src/struphy/pic/pushing/kernels/push_bxu_Hdiv/__init__.py
index ce4a76f60..0fe720316 100644
--- a/src/struphy/pic/pushing/kernels/push_bxu_Hdiv/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_bxu_Hdiv/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_bxu_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_bxu_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_bxu_Hdiv_pauli/__init__.py b/src/struphy/pic/pushing/kernels/push_bxu_Hdiv_pauli/__init__.py
index 124a2aa03..588109995 100644
--- a/src/struphy/pic/pushing/kernels/push_bxu_Hdiv_pauli/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_bxu_Hdiv_pauli/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_bxu_Hdiv_pauli = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_bxu_Hdiv_pauli = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_deterministic_diffusion_stage/__init__.py b/src/struphy/pic/pushing/kernels/push_deterministic_diffusion_stage/__init__.py
index 28b8bc04a..b1675edb1 100644
--- a/src/struphy/pic/pushing/kernels/push_deterministic_diffusion_stage/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_deterministic_diffusion_stage/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_deterministic_diffusion_stage = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_deterministic_diffusion_stage = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_eta_stage/__init__.py b/src/struphy/pic/pushing/kernels/push_eta_stage/__init__.py
index c15cc7df7..8eb94d296 100644
--- a/src/struphy/pic/pushing/kernels/push_eta_stage/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_eta_stage/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_eta_stage = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_eta_stage = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_1st_order/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_1st_order/__init__.py
index 679af23ac..f1e64f813 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_1st_order/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_1st_order/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_Bstar_discrete_gradient_1st_order = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_Bstar_discrete_gradient_1st_order = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_1st_order_newton/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_1st_order_newton/__init__.py
index c2353c0b2..0b34ce8e5 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_1st_order_newton/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_1st_order_newton/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_Bstar_discrete_gradient_1st_order_newton = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_Bstar_discrete_gradient_1st_order_newton = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_2nd_order/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_2nd_order/__init__.py
index 36ccd7504..b29a2e6d4 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_2nd_order/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_Bstar_discrete_gradient_2nd_order/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_Bstar_discrete_gradient_2nd_order = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_Bstar_discrete_gradient_2nd_order = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_gc_Bstar_explicit_multistage/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_Bstar_explicit_multistage/__init__.py
index 6c3cc2553..93f13d700 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_Bstar_explicit_multistage/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_Bstar_explicit_multistage/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_Bstar_explicit_multistage = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_Bstar_explicit_multistage = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_gc_bxEstar_dg_1st_order_newton/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_bxEstar_dg_1st_order_newton/__init__.py
index 9d338fb59..cde9382a0 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_bxEstar_dg_1st_order_newton/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_bxEstar_dg_1st_order_newton/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_bxEstar_dg_1st_order_newton = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_bxEstar_dg_1st_order_newton = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_gc_bxEstar_discrete_gradient_1st_order/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_bxEstar_discrete_gradient_1st_order/__init__.py
index b448e012c..cda6756a8 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_bxEstar_discrete_gradient_1st_order/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_bxEstar_discrete_gradient_1st_order/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_bxEstar_discrete_gradient_1st_order = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_bxEstar_discrete_gradient_1st_order = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_gc_bxEstar_discrete_gradient_2nd_order/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_bxEstar_discrete_gradient_2nd_order/__init__.py
index 9f85ee929..2f33964a7 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_bxEstar_discrete_gradient_2nd_order/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_bxEstar_discrete_gradient_2nd_order/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_bxEstar_discrete_gradient_2nd_order = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_bxEstar_discrete_gradient_2nd_order = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_gc_bxEstar_explicit_multistage/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_bxEstar_explicit_multistage/__init__.py
index 3b9218ad5..3b1385b20 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_bxEstar_explicit_multistage/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_bxEstar_explicit_multistage/__init__.py
@@ -4,4 +4,9 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_bxEstar_explicit_multistage = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_bxEstar_explicit_multistage = Kernel.from_folder(
+    __name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS}
+)
diff --git a/src/struphy/pic/pushing/kernels/push_gc_cc_J1_H1vec/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_cc_J1_H1vec/__init__.py
index 199881490..074ba4283 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_cc_J1_H1vec/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_cc_J1_H1vec/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_cc_J1_H1vec = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_cc_J1_H1vec = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_gc_cc_J1_Hcurl/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_cc_J1_Hcurl/__init__.py
index 05256db56..ff6c49b88 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_cc_J1_Hcurl/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_cc_J1_Hcurl/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_cc_J1_Hcurl = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_cc_J1_Hcurl = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_gc_cc_J1_Hdiv/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_cc_J1_Hdiv/__init__.py
index 090ff3eef..11ad60e87 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_cc_J1_Hdiv/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_cc_J1_Hdiv/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_cc_J1_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_cc_J1_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_gc_cc_J2_dg_Hdiv/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_cc_J2_dg_Hdiv/__init__.py
index 05770fb3a..c8b8cdcca 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_cc_J2_dg_Hdiv/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_cc_J2_dg_Hdiv/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_cc_J2_dg_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (1,)
+
+push_gc_cc_J2_dg_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_gc_cc_J2_dg_init_Hdiv/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_cc_J2_dg_init_Hdiv/__init__.py
index 064befbc0..314151685 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_cc_J2_dg_init_Hdiv/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_cc_J2_dg_init_Hdiv/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_cc_J2_dg_init_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (1,)
+
+push_gc_cc_J2_dg_init_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_gc_cc_J2_stage_H1vec/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_cc_J2_stage_H1vec/__init__.py
index 91054e0bf..1e4d4fdf4 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_cc_J2_stage_H1vec/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_cc_J2_stage_H1vec/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_cc_J2_stage_H1vec = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_cc_J2_stage_H1vec = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_gc_cc_J2_stage_Hdiv/__init__.py b/src/struphy/pic/pushing/kernels/push_gc_cc_J2_stage_Hdiv/__init__.py
index 8ebfc7b16..ff9573923 100644
--- a/src/struphy/pic/pushing/kernels/push_gc_cc_J2_stage_Hdiv/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_gc_cc_J2_stage_Hdiv/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_gc_cc_J2_stage_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_gc_cc_J2_stage_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_pc_GXu/__init__.py b/src/struphy/pic/pushing/kernels/push_pc_GXu/__init__.py
index bd6eb9d4e..527345f66 100644
--- a/src/struphy/pic/pushing/kernels/push_pc_GXu/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_pc_GXu/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_pc_GXu = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_pc_GXu = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_pc_GXu_full/__init__.py b/src/struphy/pic/pushing/kernels/push_pc_GXu_full/__init__.py
index 05b5e8a31..e99a58374 100644
--- a/src/struphy/pic/pushing/kernels/push_pc_GXu_full/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_pc_GXu_full/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_pc_GXu_full = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_pc_GXu_full = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_pc_eta_stage_H1vec/__init__.py b/src/struphy/pic/pushing/kernels/push_pc_eta_stage_H1vec/__init__.py
index 4115028d7..585ffd144 100644
--- a/src/struphy/pic/pushing/kernels/push_pc_eta_stage_H1vec/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_pc_eta_stage_H1vec/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_pc_eta_stage_H1vec = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_pc_eta_stage_H1vec = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_pc_eta_stage_Hcurl/__init__.py b/src/struphy/pic/pushing/kernels/push_pc_eta_stage_Hcurl/__init__.py
index 19d494201..2b16c329a 100644
--- a/src/struphy/pic/pushing/kernels/push_pc_eta_stage_Hcurl/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_pc_eta_stage_Hcurl/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_pc_eta_stage_Hcurl = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_pc_eta_stage_Hcurl = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_pc_eta_stage_Hdiv/__init__.py b/src/struphy/pic/pushing/kernels/push_pc_eta_stage_Hdiv/__init__.py
index d0b559420..cc12613ea 100644
--- a/src/struphy/pic/pushing/kernels/push_pc_eta_stage_Hdiv/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_pc_eta_stage_Hdiv/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_pc_eta_stage_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_pc_eta_stage_Hdiv = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_random_diffusion_stage/__init__.py b/src/struphy/pic/pushing/kernels/push_random_diffusion_stage/__init__.py
index ead4c2b54..c1664863f 100644
--- a/src/struphy/pic/pushing/kernels/push_random_diffusion_stage/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_random_diffusion_stage/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_random_diffusion_stage = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_random_diffusion_stage = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_v_sph_pressure/__init__.py b/src/struphy/pic/pushing/kernels/push_v_sph_pressure/__init__.py
index 59a384356..07cbec3cd 100644
--- a/src/struphy/pic/pushing/kernels/push_v_sph_pressure/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_v_sph_pressure/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_v_sph_pressure = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_v_sph_pressure = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_v_sph_pressure_ideal_gas/__init__.py b/src/struphy/pic/pushing/kernels/push_v_sph_pressure_ideal_gas/__init__.py
index 0ac91045b..94009992e 100644
--- a/src/struphy/pic/pushing/kernels/push_v_sph_pressure_ideal_gas/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_v_sph_pressure_ideal_gas/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_v_sph_pressure_ideal_gas = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_v_sph_pressure_ideal_gas = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_v_viscosity/__init__.py b/src/struphy/pic/pushing/kernels/push_v_viscosity/__init__.py
index 9f97cbd54..3af102414 100644
--- a/src/struphy/pic/pushing/kernels/push_v_viscosity/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_v_viscosity/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_v_viscosity = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_v_viscosity = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_v_with_efield/__init__.py b/src/struphy/pic/pushing/kernels/push_v_with_efield/__init__.py
index 32e8138c1..d4fd5d948 100644
--- a/src/struphy/pic/pushing/kernels/push_v_with_efield/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_v_with_efield/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_v_with_efield = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_v_with_efield = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_vxb_analytic/__init__.py b/src/struphy/pic/pushing/kernels/push_vxb_analytic/__init__.py
index 7cf7779c5..1d1d16363 100644
--- a/src/struphy/pic/pushing/kernels/push_vxb_analytic/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_vxb_analytic/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_vxb_analytic = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_vxb_analytic = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_vxb_implicit/__init__.py b/src/struphy/pic/pushing/kernels/push_vxb_implicit/__init__.py
index cd0b8f3ff..ccbc03b2b 100644
--- a/src/struphy/pic/pushing/kernels/push_vxb_implicit/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_vxb_implicit/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_vxb_implicit = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_vxb_implicit = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/push_weights_with_efield_lin_va/__init__.py b/src/struphy/pic/pushing/kernels/push_weights_with_efield_lin_va/__init__.py
index 9cd7f9228..1f145291a 100644
--- a/src/struphy/pic/pushing/kernels/push_weights_with_efield_lin_va/__init__.py
+++ b/src/struphy/pic/pushing/kernels/push_weights_with_efield_lin_va/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-push_weights_with_efield_lin_va = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+push_weights_with_efield_lin_va = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/reflect/__init__.py b/src/struphy/pic/pushing/kernels/reflect/__init__.py
index 31c85772d..35480bdb7 100644
--- a/src/struphy/pic/pushing/kernels/reflect/__init__.py
+++ b/src/struphy/pic/pushing/kernels/reflect/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-reflect = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: markers
+OUTPUTS = (0,)
+
+reflect = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/sph_isotherm_kappa/__init__.py b/src/struphy/pic/pushing/kernels/sph_isotherm_kappa/__init__.py
index 8aee82be1..62653c52c 100644
--- a/src/struphy/pic/pushing/kernels/sph_isotherm_kappa/__init__.py
+++ b/src/struphy/pic/pushing/kernels/sph_isotherm_kappa/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-sph_isotherm_kappa = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+sph_isotherm_kappa = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/sph_mean_velocity_coeffs/__init__.py b/src/struphy/pic/pushing/kernels/sph_mean_velocity_coeffs/__init__.py
index 064a0f411..6da31fac3 100644
--- a/src/struphy/pic/pushing/kernels/sph_mean_velocity_coeffs/__init__.py
+++ b/src/struphy/pic/pushing/kernels/sph_mean_velocity_coeffs/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-sph_mean_velocity_coeffs = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+sph_mean_velocity_coeffs = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/sph_pressure_coeffs/__init__.py b/src/struphy/pic/pushing/kernels/sph_pressure_coeffs/__init__.py
index 23c1fc506..8889b956e 100644
--- a/src/struphy/pic/pushing/kernels/sph_pressure_coeffs/__init__.py
+++ b/src/struphy/pic/pushing/kernels/sph_pressure_coeffs/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-sph_pressure_coeffs = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+sph_pressure_coeffs = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/sph_viscosity_tensor/__init__.py b/src/struphy/pic/pushing/kernels/sph_viscosity_tensor/__init__.py
index e9af1c7b7..fc5f3ef3a 100644
--- a/src/struphy/pic/pushing/kernels/sph_viscosity_tensor/__init__.py
+++ b/src/struphy/pic/pushing/kernels/sph_viscosity_tensor/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-sph_viscosity_tensor = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+sph_viscosity_tensor = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/pushing/kernels/unit_b_1form/__init__.py b/src/struphy/pic/pushing/kernels/unit_b_1form/__init__.py
index a39be1798..320ca9a67 100644
--- a/src/struphy/pic/pushing/kernels/unit_b_1form/__init__.py
+++ b/src/struphy/pic/pushing/kernels/unit_b_1form/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-unit_b_1form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: args_markers
+OUTPUTS = (2,)
+
+unit_b_1form = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/sph/kernels/box_based_evaluation_flat/__init__.py b/src/struphy/pic/sph/kernels/box_based_evaluation_flat/__init__.py
index 0b319a513..b0fbb0dc6 100644
--- a/src/struphy/pic/sph/kernels/box_based_evaluation_flat/__init__.py
+++ b/src/struphy/pic/sph/kernels/box_based_evaluation_flat/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-box_based_evaluation_flat = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: out
+OUTPUTS = (19,)
+
+box_based_evaluation_flat = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/sph/kernels/box_based_evaluation_meshgrid/__init__.py b/src/struphy/pic/sph/kernels/box_based_evaluation_meshgrid/__init__.py
index f2e706de3..e989365b1 100644
--- a/src/struphy/pic/sph/kernels/box_based_evaluation_meshgrid/__init__.py
+++ b/src/struphy/pic/sph/kernels/box_based_evaluation_meshgrid/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-box_based_evaluation_meshgrid = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: out
+OUTPUTS = (19,)
+
+box_based_evaluation_meshgrid = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/sph/kernels/naive_evaluation_flat/__init__.py b/src/struphy/pic/sph/kernels/naive_evaluation_flat/__init__.py
index 9b026904b..c8cd8dfca 100644
--- a/src/struphy/pic/sph/kernels/naive_evaluation_flat/__init__.py
+++ b/src/struphy/pic/sph/kernels/naive_evaluation_flat/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-naive_evaluation_flat = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: out
+OUTPUTS = (13,)
+
+naive_evaluation_flat = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/sph/kernels/naive_evaluation_meshgrid/__init__.py b/src/struphy/pic/sph/kernels/naive_evaluation_meshgrid/__init__.py
index 60de44768..10fa42ba8 100644
--- a/src/struphy/pic/sph/kernels/naive_evaluation_meshgrid/__init__.py
+++ b/src/struphy/pic/sph/kernels/naive_evaluation_meshgrid/__init__.py
@@ -4,4 +4,7 @@
 
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
-naive_evaluation_meshgrid = Kernel.from_folder(__name__, structs=CUDA_STRUCTS)
+# the arguments the kernel writes to, by position: out
+OUTPUTS = (13,)
+
+naive_evaluation_meshgrid = Kernel.from_folder(__name__, structs=CUDA_STRUCTS, host_options={"outputs": OUTPUTS})
diff --git a/src/struphy/pic/tests/cuda_parity_cases.py b/src/struphy/pic/tests/cuda_parity_cases.py
index e2ce0a5d5..024318ff8 100644
--- a/src/struphy/pic/tests/cuda_parity_cases.py
+++ b/src/struphy/pic/tests/cuda_parity_cases.py
@@ -1,8 +1,9 @@
 """Inputs for the pyccel/CUDA parity tests: for every kernel with a CUDA version, the cases it is checked on.
 
 Test-only: ``test_cuda_parity.py`` (on a GPU) and ``test_cuda_emulation.py`` (without one) run each case on both
-kernel versions and compare every array. ``build(case)`` returns the kernel's positional arguments on the active
-backend; a kernel that is ported to CUDA gets an entry here.
+kernel versions and compare the arrays of the arguments the kernel declares as outputs (``OUTPUTS`` in its folder's
+``__init__.py``); the other arrays must stay unchanged. ``build(case)`` returns the kernel's positional arguments on
+the active backend; a kernel that is ported to CUDA gets an entry here.
 """
 
 from collections.abc import Callable, Sequence
@@ -14,6 +15,8 @@
 
 from struphy.geometry.base import inside_logical_cube
 from struphy.geometry.domains import Cuboid
+from struphy.kernel_arguments.local_projectors_args_cuda import CudaLocalProjectorsArguments
+from struphy.kernel_arguments.pusher_args_cuda import CudaDerhamArguments, CudaDomainArguments, CudaMarkerArguments
 from struphy.ode.utils import ButcherTableau
 from struphy.pic.tests.kernel_test_args import (
     BOUNDARY_CONDITIONS,
@@ -45,6 +48,32 @@ class ParityCases:
     """Launch size, if it is not one thread per row of the first array."""
 
 
+# pyccel argument class name -> the fields of its CUDA struct, the data of an argument object (the pyccel scratch
+# arrays bn1, ..., bd3 of DerhamArguments are not fields)
+STRUCT_FIELDS = {
+    cls.__name__.removeprefix("Cuda"): tuple(field.name for field in cls.struct.fields)
+    for cls in (CudaMarkerArguments, CudaDerhamArguments, CudaDomainArguments, CudaLocalProjectorsArguments)
+}
+
+
+def argument_arrays(args, indices=None):
+    """Host copies of the arrays among `args` (or among the arguments `indices`), by name.
+
+    An array argument is named ``"argument <i>"``; a pyccel argument object contributes the arrays among its struct
+    fields, ``"argument <i>.<field>"`` (the names of ``cunumpy.kernel_testing.assert_kernels_agree``).
+    """
+    found = {}
+    for i in range(len(args)) if indices is None else (i % len(args) for i in indices):
+        value = args[i]
+        if isinstance(value, np.ndarray):
+            found[f"argument {i}"] = value.copy()
+        for field in STRUCT_FIELDS.get(type(value).__name__, ()):
+            item = getattr(value, field, None)
+            if isinstance(item, np.ndarray):
+                found[f"argument {i}.{field}"] = item.copy()
+    return found
+
+
 # ---------------------------------------------------------------- pushers
 
 
diff --git a/src/struphy/pic/tests/test_cuda_emulation.py b/src/struphy/pic/tests/test_cuda_emulation.py
index 793bb9525..9ecec905f 100644
--- a/src/struphy/pic/tests/test_cuda_emulation.py
+++ b/src/struphy/pic/tests/test_cuda_emulation.py
@@ -9,50 +9,36 @@
 from cunumpy.kernel_testing import emulation_compiler
 
 from struphy.geometry.tests import spline_mapping_cases
-from struphy.kernel_arguments.pusher_args_cuda import CudaDerhamArguments, CudaDomainArguments, CudaMarkerArguments
 from struphy.pic.tests.cuda_emulation import emulate_struct_kernel
-from struphy.pic.tests.cuda_parity_cases import PARITY_CASES
+from struphy.pic.tests.cuda_parity_cases import PARITY_CASES, argument_arrays
 from struphy.pic.tests.kernel_test_args import N_GEOMETRY_DOMAINS
 from struphy.pic.tests.test_cuda_parity import CUDA_KERNELS
 
-# pyccel argument class name -> its CUDA version
-CUDA_CLASSES = {
-    cls.__name__.removeprefix("Cuda"): cls for cls in (CudaMarkerArguments, CudaDerhamArguments, CudaDomainArguments)
-}
-
 requires_compiler = pytest.mark.skipif(emulation_compiler() is None, reason="no C++ compiler")
 
 
-def arrays(args):
-    """The arrays among the arguments and inside the argument objects, as host copies."""
-    out = []
-    for a in args:
-        if isinstance(a, np.ndarray):
-            out.append(a.copy())
-        elif type(a).__name__ in CUDA_CLASSES:  # a pyccel argument object: the fields of its CUDA struct
-            fields = (getattr(a, f.name, None) for f in CUDA_CLASSES[type(a).__name__].struct.fields)
-            out += [f.copy() for f in fields if isinstance(f, np.ndarray)]
-    return out
-
-
 @requires_compiler
 @pytest.mark.parametrize("name", list(PARITY_CASES))
 def test_emulated_parity(name):
+    """The emulated CUDA kernel writes the declared outputs like the pyccel kernel, and nothing else."""
     kernel, spec = CUDA_KERNELS[name], PARITY_CASES[name]
+    outputs = kernel.host_kernel.outputs
     with xp.use_backend("numpy"):
         for case in spec.cases:
             host_args = spec.build(case)
             emulated_args = spec.build(case)
-            before = arrays(host_args)
+            before = argument_arrays(emulated_args)
             kernel(*host_args)
             n_threads = spec.n_threads(emulated_args) if spec.n_threads else None
             emulate_struct_kernel(kernel.cuda_kernel, *emulated_args, n_threads=n_threads)
-            host, emulated = arrays(host_args), arrays(emulated_args)
-            assert any(not np.array_equal(a, b) for a, b in zip(before, host)), (
-                f"{name}{case}: the kernel changed nothing"
-            )
-            for h, e in zip(host, emulated):
-                np.testing.assert_allclose(e, h, rtol=spec.rtol, atol=spec.atol, err_msg=f"{name}{case}")
+            host, emulated = argument_arrays(host_args, outputs), argument_arrays(emulated_args, outputs)
+            for key, value in argument_arrays(emulated_args).items():
+                if key not in emulated:
+                    np.testing.assert_array_equal(value, before[key], err_msg=f"{name}{case}: {key} is not an output")
+            for key in host:
+                np.testing.assert_allclose(
+                    emulated[key], host[key], rtol=spec.rtol, atol=spec.atol, err_msg=f"{name}{case}: {key}"
+                )
 
 
 @requires_compiler
diff --git a/src/struphy/pic/tests/test_cuda_parity.py b/src/struphy/pic/tests/test_cuda_parity.py
index 75d53245f..7d5011145 100644
--- a/src/struphy/pic/tests/test_cuda_parity.py
+++ b/src/struphy/pic/tests/test_cuda_parity.py
@@ -1,4 +1,9 @@
-"""Parity of every CUDA kernel with its pyccel kernel, on the cases of ``cuda_parity_cases.PARITY_CASES``."""
+"""Parity of every CUDA kernel with its pyccel kernel, on the cases of ``cuda_parity_cases.PARITY_CASES``.
+
+The parity test compares the arguments each kernel declares as outputs (``OUTPUTS`` in its folder's
+``__init__.py``, passed to cunumpy as ``host_options={"outputs": OUTPUTS}``); ``test_declared_outputs`` checks the
+declarations on the NumPy backend.
+"""
 
 import importlib
 
@@ -15,7 +20,7 @@
 from struphy.pic.pushing.kernels.push_vxb_analytic import push_vxb_analytic
 from struphy.pic.pushing.kernels.push_vxb_implicit import push_vxb_implicit
 from struphy.pic.pushing.kernels.push_weights_with_efield_lin_va import push_weights_with_efield_lin_va
-from struphy.pic.tests.cuda_parity_cases import PARITY_CASES
+from struphy.pic.tests.cuda_parity_cases import PARITY_CASES, argument_arrays
 from struphy.utils.cuda_arguments import CUDA_STRUCTS
 
 # all kernel packages (one folder per kernel), for tests that go through every kernel
@@ -30,8 +35,13 @@
     "struphy.feec.local_projectors.kernels",
 )
 CATALOGS = {package: KernelCatalog.from_package(package, structs=CUDA_STRUCTS) for package in PACKAGES}
+# the kernels as their folders declare them (with OUTPUTS, which from_package does not know), by package
+DECLARED = {
+    package: KernelCatalog({name: getattr(importlib.import_module(f"{package}.{name}"), name) for name in catalog})
+    for package, catalog in CATALOGS.items()
+}
 # every kernel with a CUDA version, by name
-CUDA_KERNELS = {name: kernel for catalog in CATALOGS.values() for name, kernel in catalog.parity_cases()}
+CUDA_KERNELS = {name: kernel for catalog in DECLARED.values() for name, kernel in catalog.parity_cases()}
 # (kernel name, case index) of every parity case, as pytest parameters
 PARITY_PARAMS = [
     pytest.param(name, index, id=f"{name}-{index}")
@@ -68,9 +78,37 @@ def test_cuda_kernels_have_parity_cases():
     assert all(len(spec.cases) > 0 for spec in PARITY_CASES.values())
 
 
+def test_kernels_declare_outputs():
+    """Every kernel folder declares the arguments its kernel writes to, by position (``OUTPUTS``)."""
+    for catalog in DECLARED.values():
+        for name, kernel in catalog.items():
+            outputs, n_params = kernel.host_kernel.outputs, len(kernel.host_parameters())
+            assert outputs, f"{name}: declare OUTPUTS in its __init__.py"
+            assert all(isinstance(i, int) and -n_params <= i < n_params for i in outputs), (name, outputs)
+
+
+@pytest.mark.parametrize("name", list(PARITY_CASES))
+def test_declared_outputs(name):
+    """On the NumPy backend, the kernel changes its declared outputs and no other argument, in every parity case."""
+    kernel, spec = CUDA_KERNELS[name], PARITY_CASES[name]
+    with xp.use_backend("numpy"):
+        for case in spec.cases:
+            args = spec.build(case)
+            before = argument_arrays(args)
+            kernel(*args)
+            outputs = argument_arrays(args, kernel.host_kernel.outputs)
+            assert any(not np.array_equal(value, before[key]) for key, value in outputs.items()), (
+                f"{name}{case}: the kernel changed none of its outputs"
+            )
+            for key, value in argument_arrays(args).items():
+                if key not in outputs:
+                    np.testing.assert_array_equal(value, before[key], err_msg=f"{name}{case}: {key} is not an output")
+
+
 @requires_cupy
 @pytest.mark.parametrize("name, index", PARITY_PARAMS)
 def test_parity(name, index):
+    """assert_kernels_agree compares the declared outputs of the kernel (``kernel.host_kernel.outputs``)."""
     check_case(name, index, assert_kernels_agree)