PyOpenCL target: Add, test overflow of large argument counts into SVM struct

Co-authored-by: Matthias Diener <[email protected]>

Author: inducer

loopy/codegen/result.py

@@ -27,6 +27,10 @@
 import islpy as isl
 
 
+if TYPE_CHECKING:
+    from loopy.codegen import CodeGenerationState
+
+
 def process_preambles(preambles: Sequence[Tuple[int, str]]) -> Sequence[str]:
     seen_preamble_tags = set()
     dedup_preambles = []
@@ -171,7 +175,8 @@ def all_code(self):
                 + "\n\n"
                 + str(self.host_program.ast))
 
-    def current_program(self, codegen_state):
+    def current_program(
+            self, codegen_state: "CodeGenerationState") -> GeneratedProgram:
         if codegen_state.is_generating_device_code:
             if self.device_programs:
                 result = self.device_programs[-1]
@@ -344,13 +349,23 @@ def generate_host_or_device_program(codegen_state, schedule_index):
 
         cur_prog = codegen_result.current_program(codegen_state)
         body_ast = cur_prog.ast
-        fdecl_ast = ast_builder.get_function_declaration(
+        fdef_preambles, fdecl_ast = ast_builder.get_function_declaration(
                 codegen_state, codegen_result, schedule_index)
 
         fdef_ast = ast_builder.get_function_definition(
                 codegen_state, codegen_result,
                 schedule_index, fdecl_ast, body_ast)
 
+        if fdef_preambles:
+            if codegen_state.is_generating_device_code:
+                codegen_result = codegen_result.copy(
+                        device_preambles=(
+                            codegen_result.device_preambles + tuple(fdef_preambles)))
+            else:
+                codegen_result = codegen_result.copy(
+                        host_preambles=(
+                            codegen_result.host_preambles + tuple(fdef_preambles)))
+
         codegen_result = codegen_result.with_new_program(
                 codegen_state,
                 cur_prog.copy(

loopy/target/__init__.py

@@ -203,7 +203,8 @@ def get_function_definition(
     def get_function_declaration(
             self, codegen_state: CodeGenerationState,
             codegen_result: CodeGenerationResult, schedule_index: int
-            ) -> ASTType:
+            ) -> Tuple[Sequence[Tuple[str, str]], ASTType]:
+        """Returns preambles and the AST for the function declaration."""
         raise NotImplementedError
 
     def generate_top_of_body(
@@ -293,14 +294,16 @@ def __str__(self):
         return ""
 
 
-class DummyHostASTBuilder(ASTBuilderBase):
+class DummyHostASTBuilder(ASTBuilderBase[None]):
     def get_function_definition(self, codegen_state, codegen_result,
             schedule_index, function_decl, function_body):
         return function_body
 
-    def get_function_declaration(self, codegen_state, codegen_result,
-            schedule_index):
-        return None
+    def get_function_declaration(
+            self, codegen_state, codegen_result,
+            schedule_index,
+            ) -> Tuple[Sequence[Tuple[str, str]], None]:
+        return [], None
 
     def get_temporary_decls(self, codegen_state, schedule_index):
         return []

loopy/target/c/__init__.py

@@ -23,7 +23,7 @@
 THE SOFTWARE.
 """
 
-from typing import cast, Tuple, Optional
+from typing import cast, Tuple, Optional, Sequence
 import re
 
 import numpy as np  # noqa
@@ -817,8 +817,10 @@ def get_function_definition(
         else:
             return Collection(result+[Line(), fbody])
 
-    def get_function_declaration(self, codegen_state: CodeGenerationState,
-            codegen_result: CodeGenerationResult, schedule_index: int) -> Generable:
+    def get_function_declaration(
+            self, codegen_state: CodeGenerationState,
+            codegen_result: CodeGenerationResult, schedule_index: int
+            ) -> Tuple[Sequence[Tuple[str, str]], Generable]:
         kernel = codegen_state.kernel
 
         assert codegen_state.kernel.linearization is not None
@@ -846,7 +848,7 @@ def get_function_declaration(self, codegen_state: CodeGenerationState,
             passed_names = [arg.name for arg in kernel.args]
             written_names = kernel.get_written_variables()
 
-        return FunctionDeclarationWrapper(
+        return [], FunctionDeclarationWrapper(
                 FunctionDeclaration(
                     name,
                     [self.arg_to_cgen_declarator(

loopy/target/cuda.py

@@ -23,10 +23,12 @@
 THE SOFTWARE.
 """
 
+from typing import Tuple, Sequence
+
 import numpy as np
 from pymbolic import var
 from pytools import memoize_method
-from cgen import Declarator, Const
+from cgen import Declarator, Const, Generable
 
 from loopy.target.c import CFamilyTarget, CFamilyASTBuilder
 from loopy.target.c.codegen.expression import ExpressionToCExpressionMapper
@@ -35,6 +37,8 @@
 from loopy.kernel.array import ArrayBase, FixedStrideArrayDimTag, VectorArrayDimTag
 from loopy.kernel.data import AddressSpace, ImageArg, ConstantArg, ArrayArg
 from loopy.kernel.function_interface import ScalarCallable
+from loopy.codegen.result import CodeGenerationResult
+from loopy.codegen import CodeGenerationState
 
 
 # {{{ vector types
@@ -316,9 +320,11 @@ def known_callables(self):
 
     # {{{ top-level codegen
 
-    def get_function_declaration(self, codegen_state, codegen_result,
-            schedule_index):
-        fdecl = super().get_function_declaration(
+    def get_function_declaration(
+            self, codegen_state: CodeGenerationState,
+            codegen_result: CodeGenerationResult, schedule_index: int
+            ) -> Tuple[Sequence[Tuple[str, str]], Generable]:
+        preambles, fdecl = super().get_function_declaration(
                 codegen_state, codegen_result, schedule_index)
 
         from loopy.target.c import FunctionDeclarationWrapper
@@ -348,7 +354,7 @@ def get_function_declaration(self, codegen_state, codegen_result,
 
             fdecl = CudaLaunchBounds(nthreads, fdecl)
 
-        return FunctionDeclarationWrapper(fdecl)
+        return preambles, FunctionDeclarationWrapper(fdecl)
 
     def preamble_generators(self):
 

loopy/target/ispc.py

@@ -24,7 +24,7 @@
 """
 
 
-from typing import cast, Tuple
+from typing import cast, Tuple, Sequence
 
 import numpy as np  # noqa
 import pymbolic.primitives as p
@@ -202,8 +202,10 @@ def get_dtype_registry(self):
 class ISPCASTBuilder(CFamilyASTBuilder):
     # {{{ top-level codegen
 
-    def get_function_declaration(self, codegen_state: CodeGenerationState,
-            codegen_result: CodeGenerationResult, schedule_index: int) -> Generable:
+    def get_function_declaration(
+            self, codegen_state: CodeGenerationState,
+            codegen_result: CodeGenerationResult, schedule_index: int
+            ) -> Tuple[Sequence[Tuple[str, str]], Generable]:
         name = codegen_result.current_program(codegen_state).name
         kernel = codegen_state.kernel
 
@@ -243,7 +245,7 @@ def get_function_declaration(self, codegen_state: CodeGenerationState,
                         arg_decls))
 
         from loopy.target.c import FunctionDeclarationWrapper
-        return FunctionDeclarationWrapper(result)
+        return [], FunctionDeclarationWrapper(result)
 
     def get_kernel_call(self, codegen_state: CodeGenerationState,
             subkernel_name: str,

loopy/target/opencl.py

@@ -23,10 +23,12 @@
 THE SOFTWARE.
 """
 
+from typing import Tuple, Sequence
+
 import numpy as np
 from pymbolic import var
 from pytools import memoize_method
-from cgen import Declarator
+from cgen import Declarator, Generable
 
 from loopy.target.c import CFamilyTarget, CFamilyASTBuilder
 from loopy.target.c.codegen.expression import ExpressionToCExpressionMapper
@@ -36,6 +38,8 @@
 from loopy.kernel.array import VectorArrayDimTag, FixedStrideArrayDimTag, ArrayBase
 from loopy.kernel.data import AddressSpace, ImageArg, ConstantArg
 from loopy.kernel.function_interface import ScalarCallable
+from loopy.codegen import CodeGenerationState
+from loopy.codegen.result import CodeGenerationResult
 
 
 # {{{ dtype registry wrappers
@@ -624,20 +628,26 @@ def preamble_generators(self):
 
     # {{{ top-level codegen
 
-    def get_function_declaration(self, codegen_state, codegen_result,
-            schedule_index):
-        fdecl = super().get_function_declaration(
+    def get_function_declaration(
+            self, codegen_state: CodeGenerationState,
+            codegen_result: CodeGenerationResult, schedule_index: int
+            ) -> Tuple[Sequence[Tuple[str, str]], Generable]:
+        preambles, fdecl = super().get_function_declaration(
                 codegen_state, codegen_result, schedule_index)
 
         from loopy.target.c import FunctionDeclarationWrapper
         assert isinstance(fdecl, FunctionDeclarationWrapper)
         if not codegen_state.is_entrypoint:
             # auxiliary kernels need not mention opencl speicific qualifiers
             # for a functions signature
-            return fdecl
+            return preambles, fdecl
 
-        fdecl = fdecl.subdecl
+        return preambles, FunctionDeclarationWrapper(
+                self._wrap_kernel_decl(codegen_state, schedule_index, fdecl.subdecl))
 
+    def _wrap_kernel_decl(
+            self, codegen_state: CodeGenerationState, schedule_index: int,
+            fdecl: Declarator) -> Declarator:
         from cgen.opencl import CLKernel, CLRequiredWorkGroupSize
         fdecl = CLKernel(fdecl)
 
@@ -654,7 +664,7 @@ def get_function_declaration(self, codegen_state, codegen_result,
 
             fdecl = CLRequiredWorkGroupSize(local_sizes, fdecl)
 
-        return FunctionDeclarationWrapper(fdecl)
+        return fdecl
 
     def generate_top_of_body(self, codegen_state):
         from loopy.kernel.data import ImageArg