diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 78e6f2434c548f61263b480a90ecf5b550cbe447..94e11a2d43ae30ba21d2c2da6213a5baaab96d14 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -58,10 +58,14 @@ jobs:
     pytest:
         name: Conda Pytest
         runs-on: ubuntu-latest
+        strategy:
+            matrix:
+                loopy-branch: [master, kernel_callables_v3-edit2]
         steps:
         -   uses: actions/checkout@v2
         -   name: "Main Script"
             run: |
+                sed -i "s/loopy.git/loopy.git@${{ matrix.loopy-branch }}/g" requirements.txt
                 CONDA_ENVIRONMENT=.test-conda-env-py3.yml
                 curl -L -O -k https://gitlab.tiker.net/inducer/ci-support/raw/master/build-and-test-py-project-within-miniconda.sh
                 . ./build-and-test-py-project-within-miniconda.sh
@@ -69,10 +73,14 @@ jobs:
     examples:
         name: Conda Examples
         runs-on: ubuntu-latest
+        strategy:
+            matrix:
+                loopy-branch: [master, kernel_callables_v3-edit2]
         steps:
         -   uses: actions/checkout@v2
         -   name: "Main Script"
             run: |
+                sed -i "s/loopy.git/loopy.git@${{ matrix.loopy-branch }}/g" requirements.txt
                 CONDA_ENVIRONMENT=.test-conda-env-py3.yml
                 curl -L -O -k https://gitlab.tiker.net/inducer/ci-support/raw/master/ci-support.sh
                 . ci-support.sh
diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index da24af23949c9fd052f3e945cb3d80cb689c073c..c5bbb7d9cce15fdf278a0bfc974ddeba319d711f 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -1,5 +1,6 @@
 Python 3 POCL:
   script:
+  - sed -i "s/loopy.git/loopy.git@$LOOPY_BRANCH/g" requirements.txt
   - export PY_EXE=python3
   - export PYOPENCL_TEST=portable:pthread
   - export EXTRA_INSTALL="pyopencl"
@@ -14,9 +15,15 @@ Python 3 POCL:
     reports:
       junit: test/pytest.xml
 
+  parallel:
+    matrix:
+      - LOOPY_BRANCH: master
+      - LOOPY_BRANCH: kernel_callables_v3-edit2
+
 
 Python 3 POCL Examples:
   script:
+  - sed -i "s/loopy.git/loopy.git@$LOOPY_BRANCH/g" requirements.txt
   - export PY_EXE=python3
   - export PYOPENCL_TEST=portable:pthread
   - export EXTRA_INSTALL="pyopencl"
@@ -28,6 +35,10 @@ Python 3 POCL Examples:
   - large-node
   except:
   - tags
+  parallel:
+    matrix:
+      - LOOPY_BRANCH: master
+      - LOOPY_BRANCH: kernel_callables_v3-edit2
 
 
 Pylint:
diff --git a/pytato/codegen.py b/pytato/codegen.py
index 9ca1419edc4cd3ab625e018fab3a4bff810220ed..f99afbd2c964fd94708e093a5ede153622b508ff 100644
--- a/pytato/codegen.py
+++ b/pytato/codegen.py
@@ -450,9 +450,9 @@ class CodeGenState:
 
         The (global) namespace
 
-    .. attribute:: kernel
+    .. attribute:: _program
 
-        The partial :class:`loopy.LoopKernel` being built.
+        The partial :class:`loopy.LoopKernel` or :class:`loopy.Program` being built.
 
     .. attribute:: results
 
@@ -465,22 +465,40 @@ class CodeGenState:
     .. automethod:: update_kernel
     """
     namespace: Mapping[str, Array]
-    _kernel: lp.LoopKernel
+    _program: Union["lp.Program", lp.LoopKernel]
     results: Dict[Array, ImplementedResult]
 
     var_name_gen: pytools.UniqueNameGenerator = dataclasses.field(init=False)
     insn_id_gen: pytools.UniqueNameGenerator = dataclasses.field(init=False)
 
     def __post_init__(self) -> None:
-        self.var_name_gen = self._kernel.get_var_name_generator()
-        self.insn_id_gen = self._kernel.get_instruction_id_generator()
+        if isinstance(self._program, lp.LoopKernel):
+            self.var_name_gen = self._program.get_var_name_generator()
+            self.insn_id_gen = self._program.get_instruction_id_generator()
+        else:
+            self.var_name_gen = self._program["_pt_kernel"].get_var_name_generator()
+            self.insn_id_gen = (
+                    self._program["_pt_kernel"].get_instruction_id_generator())
+
+    @property
+    def program(self) -> Union["lp.Program", lp.LoopKernel]:
+        return self._program
 
     @property
     def kernel(self) -> lp.LoopKernel:
-        return self._kernel
+        """
+        Returns the entry kernel of the loopy program being built.
+        """
+        if isinstance(self._program, lp.LoopKernel):
+            return self._program
+        else:
+            return self._program["_pt_kernel"]
 
     def update_kernel(self, kernel: lp.LoopKernel) -> None:
-        self._kernel = kernel
+        if isinstance(self._program, lp.LoopKernel):
+            self._program = kernel
+        else:
+            self._program = self._program.with_kernel(kernel)
 
 
 class CodeGenMapper(Mapper):
@@ -857,12 +875,13 @@ def normalize_outputs(result: Union[Array, DictOfNamedArrays,
 def get_initial_codegen_state(namespace: Namespace, target: Target,
         options: Optional[lp.Options]) -> CodeGenState:
     kernel = lp.make_kernel("{:}", [],
+            name="_pt_kernel",
             target=target.get_loopy_target(),
             options=options,
             lang_version=lp.MOST_RECENT_LANGUAGE_VERSION)
 
     return CodeGenState(namespace=namespace,
-            _kernel=kernel,
+            _program=kernel,
             results=dict())
 
 
@@ -919,7 +938,8 @@ def generate_loopy(result: Union[Array, DictOfNamedArrays, Dict[str, Array]],
     :param result: Outputs of the computation.
     :param target: Code generation target.
     :param options: Code generation options for the kernel.
-    :returns: A wrapped generated :mod:`loopy` kernel
+    :returns: A :class:`pytato.program.BoundProgram` wrapping the generated
+        :mod:`loopy` program.
     """
     orig_outputs: DictOfNamedArrays = normalize_outputs(result)
     del result
@@ -955,5 +975,5 @@ def generate_loopy(result: Union[Array, DictOfNamedArrays, Dict[str, Array]],
         state.results[expr] = StoredResult(name, expr.ndim, frozenset([insn_id]))
 
     return target.bind_program(
-            program=state.kernel,
+            program=state.program,
             bound_arguments=preproc_result.bound_arguments)
diff --git a/pytato/program.py b/pytato/program.py
index 2894332cc8461d49441bc888500068ea3d7d18b0..408ef152e82dcaadf80de6a51d709798ceec9fa6 100644
--- a/pytato/program.py
+++ b/pytato/program.py
@@ -34,7 +34,7 @@ Generated Executable Programs
 
 from dataclasses import dataclass
 import typing
-from typing import Any, Mapping, Optional
+from typing import Any, Mapping, Optional, Union
 
 import loopy
 
@@ -52,7 +52,7 @@ class BoundProgram:
 
     .. attribute:: program
 
-        The underlying :class:`loopy.LoopKernel`.
+        The underlying :class:`loopy.Program`.
 
     .. attribute:: target
 
@@ -65,13 +65,20 @@ class BoundProgram:
     .. automethod:: __call__
     """
 
-    program: "loopy.LoopKernel"
+    program: Union["loopy.LoopKernel", "loopy.Program"]
     bound_arguments: Mapping[str, Any]
     target: "pytato.target.Target"
 
     def __call__(self, *args: Any, **kwargs: Any) -> Any:
         raise NotImplementedError
 
+    @property
+    def kernel(self) -> "loopy.LoopKernel":
+        if isinstance(self.program, loopy.LoopKernel):
+            return self.program
+        else:
+            return self.program["_pt_kernel"]
+
 
 @dataclass(init=True, repr=False, eq=False)
 class BoundPyOpenCLProgram(BoundProgram):
@@ -92,6 +99,9 @@ class BoundPyOpenCLProgram(BoundProgram):
 
         updated_kwargs = dict(self.bound_arguments)
         updated_kwargs.update(kwargs)
+        if not isinstance(self. program, loopy.LoopKernel):
+            updated_kwargs.setdefault("entrypoint", "_pt_kernel")
+
         return self.program(self.queue, *args, **updated_kwargs)
 
 # vim: foldmethod=marker
diff --git a/pytato/target.py b/pytato/target.py
index 72e819b66e6390af0fb2e9f14a875e5ca39e63b8..162591f9ca47a888a16d12700069356881eba0cf 100644
--- a/pytato/target.py
+++ b/pytato/target.py
@@ -32,7 +32,7 @@ Code Generation Targets
 .. autoclass:: PyOpenCLTarget
 """
 
-from typing import Any, Mapping, Optional
+from typing import Any, Mapping, Optional, Union
 
 from pytato.program import BoundProgram, BoundPyOpenCLProgram
 
@@ -51,11 +51,11 @@ class Target:
         """Return the corresponding :mod:`loopy` target."""
         raise NotImplementedError
 
-    def bind_program(self, program: "loopy.LoopKernel",
+    def bind_program(self, program: Union["loopy.Program", "loopy.LoopKernel"],
             bound_arguments: Mapping[str, Any]) -> BoundProgram:
         """Create a :class:`pytato.program.BoundProgram` for this code generation target.
 
-        :param program: the :mod:`loopy` kernel
+        :param program: the :mod:`loopy` program
         :param bound_arguments: a mapping from argument names to outputs
         """
         raise NotImplementedError
@@ -79,7 +79,7 @@ class PyOpenCLTarget(Target):
             device = self.queue.device
         return lp.PyOpenCLTarget(device)
 
-    def bind_program(self, program: "loopy.LoopKernel",
+    def bind_program(self, program: Union["loopy.Program", "loopy.LoopKernel"],
             bound_arguments: Mapping[str, Any]) -> BoundProgram:
         return BoundPyOpenCLProgram(program=program,
                 queue=self.queue,
diff --git a/requirements.txt b/requirements.txt
index 5af50dd7a2258fb8d2d198e7ace09fb5bbff5022..6cbbc093945ba48d80eb1276e97fc411a21f1c57 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,2 +1,3 @@
 git+https://github.com/inducer/pytools.git#egg=pytools >= 2021.1
+git+https://github.com/inducer/genpy.git#egg=genpy
 git+https://github.com/inducer/loopy.git#egg=loopy
diff --git a/test/test_codegen.py b/test/test_codegen.py
index a537b9665b287e2ecffd972f008f30de6bf2c874..93a01c27acd94d06a128139740495acade15c89a 100755
--- a/test/test_codegen.py
+++ b/test/test_codegen.py
@@ -439,7 +439,7 @@ def test_dict_of_named_array_codegen_avoids_recomputation():
 
     yz = pt.DictOfNamedArrays({"y": y, "z": z})
 
-    knl = pt.generate_loopy(yz).program
+    knl = pt.generate_loopy(yz).kernel
     assert ("y" in knl.id_to_insn["z_store"].read_dependency_names())
 
 
@@ -470,7 +470,7 @@ def test_only_deps_as_knl_args():
     y = pt.make_placeholder(ns, name="y", shape=(10, 4), dtype=float)  # noqa:F841
 
     z = 2*x
-    knl = pt.generate_loopy(z).program
+    knl = pt.generate_loopy(z).kernel
 
     assert "x" in knl.arg_dict
     assert "y" not in knl.arg_dict
diff --git a/test/test_pytato.py b/test/test_pytato.py
index 77050e101eb94e4fa5c4baf711ff526495209af7..6804b0ac5170dd95ee7f9e914575ca326c15c1e0 100755
--- a/test/test_pytato.py
+++ b/test/test_pytato.py
@@ -119,7 +119,7 @@ def test_make_placeholder_noname():
     x = pt.make_placeholder(ns, shape=(10, 4), dtype=float)
     y = 2*x
 
-    knl = pt.generate_loopy(y).program
+    knl = pt.generate_loopy(y).kernel
 
     assert x.name in knl.arg_dict
     assert x.name in knl.get_read_variables()
@@ -132,7 +132,7 @@ def test_zero_length_arrays():
 
     assert y.shape == (0, 4)
 
-    knl = pt.generate_loopy(y).program
+    knl = pt.generate_loopy(y).kernel
     assert all(dom.is_empty() for dom in knl.domains if dom.total_dim() != 0)