Respect kernel.arguments in non-executable targets' lowered code

loopy/target/__init__.py

@@ -49,6 +49,7 @@
 
 from typing import (Any, Tuple, Generic, TypeVar, Sequence, ClassVar, Optional,
         TYPE_CHECKING)
+import abc
 
 if TYPE_CHECKING:
     from loopy.typing import ExpressionT
@@ -159,6 +160,17 @@ def get_kernel_executor(self, kernel, *args, **kwargs):
         """
         raise NotImplementedError()
 
+    @abc.abstractproperty
+    def single_subkernel_is_entrypoint(self) -> bool:
+        r"""
+        Returns *True* if *self* does NOT support generating code for
+        linearized kernels with more than one
+        :class:`~loopy.schedule.CallKernel`\ s. This guarantees the
+        :class:`~loopy.schedule.CallKernel` for which we generate code is the
+        entrypoint kernel. This also allows the target to skip the invoker
+        level code.
+        """
+
 
 class ASTBuilderBase(Generic[ASTType]):
     """An interface for generating (host or device) ASTs.

loopy/target/c/__init__.py

@@ -849,8 +849,15 @@ def get_function_declaration(
 
             # subkernel launches occur only as part of entrypoint kernels for now
             from loopy.schedule.tools import get_subkernel_arg_info
+            from loopy.kernel.tools import get_subkernels
             skai = get_subkernel_arg_info(kernel, subkernel_name)
-            passed_names = skai.passed_names
+            if (self.target.single_subkernel_is_entrypoint
+                    and len(get_subkernels(kernel)) > 1):
+                raise LoopyError(f"Kernel '{kernel.name}' has more than one"
+                                 f" subkernel, not allowed in {self.target}.")
+            passed_names = (skai.passed_names
+                            if not self.target.single_subkernel_is_entrypoint
+                            else [arg.name for arg in kernel.args])
             written_names = skai.written_names
         else:
             name = Value("static void", name)
@@ -1342,6 +1349,10 @@ def get_dtype_registry(self):
         fill_registry_with_c99_complex_types(result)
         return DTypeRegistryWrapper(result)
 
+    @property
+    def single_subkernel_is_entrypoint(self) -> bool:
+        return True
+
 
 class CASTBuilder(CFamilyASTBuilder):
     def preamble_generators(self):
@@ -1385,6 +1396,10 @@ def get_host_ast_builder(self):
         # enable host code generation
         return CFamilyASTBuilder(self)
 
+    @property
+    def single_subkernel_is_entrypoint(self) -> bool:
+        return False
+
 # }}}
 
 

loopy/target/cuda.py

@@ -255,6 +255,10 @@ def vector_dtype(self, base, count):
 
     # }}}
 
+    @property
+    def single_subkernel_is_entrypoint(self) -> bool:
+        return False
+
 # }}}
 
 

loopy/target/ispc.py

@@ -198,6 +198,10 @@ def get_dtype_registry(self):
 
     # }}}
 
+    @property
+    def single_subkernel_is_entrypoint(self) -> bool:
+        return True
+
 
 class ISPCASTBuilder(CFamilyASTBuilder):
     # {{{ top-level codegen
@@ -222,7 +226,9 @@ def get_function_declaration(
             # subkernel launches occur only as part of entrypoint kernels for now
             from loopy.schedule.tools import get_subkernel_arg_info
             skai = get_subkernel_arg_info(codegen_state.kernel, subkernel_name)
-            passed_names = skai.passed_names
+            passed_names = ([arg.name for arg in kernel.args]
+                            if self.target.single_subkernel_is_entrypoint
+                            else skai.passed_names)
             written_names = skai.written_names
         else:
             passed_names = [arg.name for arg in kernel.args]
@@ -263,7 +269,8 @@ def get_kernel_call(self, codegen_state: CodeGenerationState,
                         "assert(programCount == (%s))"
                         % ecm(lsize[0], PREC_NONE)))
 
-        if codegen_state.is_entrypoint:
+        if (codegen_state.is_entrypoint and
+                self.target.single_subkernel_is_entrypoint):
             # subkernel launches occur only as part of entrypoint kernels for now
             from loopy.schedule.tools import get_subkernel_arg_info
             skai = get_subkernel_arg_info(codegen_state.kernel, subkernel_name)

loopy/target/opencl.py

@@ -598,6 +598,10 @@ def is_vector_dtype(self, dtype):
     def vector_dtype(self, base, count):
         return NumpyType(vec.types[base.numpy_dtype, count])
 
+    @property
+    def single_subkernel_is_entrypoint(self) -> bool:
+        return False
+
 # }}}
 
 

loopy/target/pyopencl.py

@@ -645,6 +645,10 @@ def with_device(self, device):
                 "stop working in 2022.", DeprecationWarning, stacklevel=2)
         return self
 
+    @property
+    def single_subkernel_is_entrypoint(self) -> bool:
+        return False
+
 # }}}
 
 

test/test_target.py

@@ -777,6 +777,24 @@ def test_passing_bajillions_of_svm_args(ctx_factory, with_gbarrier):
         assert (res[f"c{iargset}"].get() == iargset * multiplier + iargset).all()
 
 
+def test_non_executable_targets_respect_args():
+    # See https://github.com/inducer/loopy/issues/648
+    t_unit = lp.make_kernel(
+        "{ : }",
+        """
+        a[0] = 1729
+        """,
+        [lp.GlobalArg("a,b,c,d,e",
+                      shape=(10,),
+                      dtype="float64")],
+        target=lp.CTarget()
+    )
+    code_str = lp.generate_code_v2(t_unit).device_code()
+
+    for var in ["b", "c", "d", "e"]:
+        assert code_str.find(f"double const *__restrict__ {var}") != -1
+
+
 if __name__ == "__main__":
     if len(sys.argv) > 1:
         exec(sys.argv[1])