def gen_composite_functional_kernel(g: NativeFunctionsGroup) -> Optional[str]: # We should only be generating these for code-generated NativeFunctions if "generated" not in g.functional.tags: return None # And we always write the kernel for a generated op in terms of a non-generated op. if g.inplace is not None and "generated" not in g.inplace.tags: target_f = g.inplace elif g.mutable is not None and "generated" not in g.mutable.tags: target_f = g.mutable else: # We should be guaranteed to have a valid inplace/mutable variant to call into. # See Note: [Mutable Ops Not Using Functionalization] raise AssertionError(str(g.functional.func)) sig = DispatcherSignature(g.functional.func) target_sig = DispatcherSignature(target_f.func) context: List[Union[Binding, Expr]] = [] clone_mutable_inputs = [] cloned_return_names = [] # We can't just directly pass all of the arguments from the functional op into the mutating op. # We need to check for which inputs to the mutating operator are mutable, # and clone those inputs first. for a_curr, a_tgt in zip( dispatcher.jit_arguments(g.functional.func), dispatcher.jit_arguments(target_f.func), ): if a_tgt.annotation is not None and a_tgt.annotation.is_write: clone_mutable_inputs.append( f"auto {a_curr.name}_clone = clone_arg({a_curr.name});" ) context.append( Expr( expr=f"{a_curr.name}_clone", type=dispatcher.argument_type(a_curr, binds=a_curr.name), ) ) # Invariant: mutable arguments on the inner mutable op are always returns on the functional op. cloned_return_names.append(f"{a_curr.name}_clone") else: context.append(dispatcher.argument(a_curr)) exprs = ", ".join([e.expr for e in translate(context, target_sig.arguments())]) out_name = "output" maybe_assign = f"auto {out_name} = " if len(target_f.func.returns) > 0 else "" inner_return_names = gather_nonaliased_inner_rets(target_f.func, out_name) ret_str = return_str( g.functional.func.returns, inner_return_names + cloned_return_names ) clone_mutable_inputs_str = "\n".join(clone_mutable_inputs) return f"""
def compute_ufunc_cuda_dtype_body( g: NativeFunctionsGroup, dtype: ScalarType, inner_loops: Dict[UfuncKey, UfunctorSignature], parent_ctx: Sequence[Binding], ) -> str: body = "using opmath_t = at::opmath_type<scalar_t>;" body += "if (false) {}\n" # for ease of codegen for config in BinaryScalarSpecializationConfigs: if config.ufunc_key not in inner_loops: continue ufunctor_sig = inner_loops[config.ufunc_key] scalar_idx = config.scalar_idx + 1 # Make a copy and at the same time widen the type (not permissible # without copy; we don't want to mutate the input argument anyway) ctx: List[Union[Expr, Binding]] = list(parent_ctx) ctx.append( Expr( expr=f"iter.scalar_value<opmath_t>({scalar_idx})", type=NamedCType(config.ctor_tensor, BaseCType(opmath_t)), ) ) ufunctor_ctor_exprs_str = ", ".join( a.expr for a in translate(ctx, ufunctor_sig.arguments().ctor) ) # NB: ufunctor must be allocated before iter.remove_operand is called, # as it relies on iter body += f"""\ else if (iter.is_cpu_scalar({scalar_idx})) {{ {ufunctor_sig.name}<scalar_t> ufunctor({ufunctor_ctor_exprs_str}); iter.remove_operand({scalar_idx}); gpu_kernel(iter, ufunctor); }}""" ufunctor_sig = inner_loops[UfuncKey.CUDAFunctor] ufunctor_ctor_exprs_str = ", ".join( a.expr for a in translate(parent_ctx, ufunctor_sig.arguments().ctor) ) body += f""" else {{ gpu_kernel(iter, {ufunctor_sig.name}<scalar_t>({ufunctor_ctor_exprs_str})); }} """ return body
def translate( bindings: Sequence[Union[Expr, Binding]], goals: Sequence[Union[NamedCType, Binding]], *, method: bool = False, allow_expensive_conversions: bool = False, ) -> List[Expr]: binding_exprs: List[Expr] = [] for b in bindings: if isinstance(b, Binding): binding_exprs.append(Expr( expr=b.name, type=b.nctype, )) else: binding_exprs.append(b) goal_ctypes: List[NamedCType] = [] for g in goals: if isinstance(g, Binding): goal_ctypes.append(g.nctype) else: goal_ctypes.append(g) # Add all the bindings to the context ctx: Dict[NamedCType, str] = {} for b in binding_exprs: ctx[b.type] = b.expr # While we're at it, do some simple forward inference, looking through # constructors. # # NB: When should you do forward inference versus backward inference? # The general idea: # # - Backward inference WHEN the goal gets smaller # - Forward inference WHEN the hypothesis gets smaller # # This helps ensure termination: backward inference starts with a goal # and tries to make it simpler and simpler until it's trivial; if the # goal can grow in size, we blow up to a really huge goal size. # Similarly, with forward inference we take hypotheses and decompose # them into simpler hypotheses; if hypotheses could expand in size, # we also have potential nontermination. (In the code below, forward # inference is only ever carried out at a single step, but you could # imagine repeated application of forward inference being profitable.) # # A good starting point in the literature for exploring more about proof # search are these lecture notes # https://www.cs.cmu.edu/~fp/courses/oregon-m10/04-focusing.pdf # # TODO: My kingdom for a pattern matcher # https://www.python.org/dev/peps/pep-0634/ # # TODO: This could get us in recomputation trouble if b.expr is nontrivial. # Fix this by implementing some sort of sharing so that if multiple # goals share the same expression, we only compute it once. This seems # to matter in practice as compiler is often unwilling to CSE nontrivial # expressions like scalar.to<scalar_t>() t = b.type if (isinstance(t, ConstRefCType) and isinstance(t.elem, OptionalCType) and isinstance(t.elem.elem, BaseCType) and str(t.elem.elem.type) == "at::Tensor"): ctx[NamedCType( t.elem.elem.name, ConstRefCType(BaseCType(tensorT)) )] = f"({b.expr}.has_value() ? *{b.expr} : at::Tensor())" if t.type == ConstRefCType(OptionalCType(BaseCType(tensorT))): ctx[NamedCType( t.name, BaseCType(optionalTensorRefT) )] = f"(({b.expr}.has_value() && (*{b.expr}).defined()) ? at::OptionalTensorRef(*{b.expr}) : at::OptionalTensorRef())" if t.type == ConstRefCType(BaseCType(scalarT)): ctx[NamedCType(t.name, BaseCType(opmath_t))] = f"({b.expr}).to<opmath_t>()" if t.type == ConstRefCType(OptionalCType(BaseCType(scalarT))): ctx[NamedCType( t.name, BaseCType(optionalScalarRefT) )] = f"({b.expr}.has_value() ? at::OptionalScalarRef(&({b.expr}.value())) : at::OptionalScalarRef())" if t.type == BaseCType(scalar_t): ctx[NamedCType( t.name, BaseCType(opmath_t))] = f"static_cast<opmath_t>({b.expr})" # [Note: ITensorListRef] if t.type == BaseCType(tensorListT): ctx[NamedCType( t.name, BaseCType(iTensorListRefT))] = f"at::ITensorListRef({b.expr})" # [Note: IOptTensorListRef] if t.type == ConstRefCType(ListCType(OptionalCType( BaseCType(tensorT)))): ctx[NamedCType(t.name, BaseCType( iOptTensorListRefT))] = f"at::IOptTensorListRef({b.expr})" # Add implicit bindings if the generated code is inside a Tensor method if method: ctx[NamedCType("self", MutRefCType( BaseCType(tensorT)))] = "const_cast<Tensor&>(*this)" ctx[NamedCType("self", ConstRefCType( BaseCType(tensorT)))] = "const_cast<Tensor&>(*this)" # This is better! Byte-for-byte compat # ctx[NamedCType("self", ConstRefCType(BaseCType(tensorT)))] = "*this" def unsat(goal: NamedCType) -> NoReturn: ctx_desc = "\n".join(f" {t.cpp_type()} {t.name}; // {e}" for t, e in ctx.items()) raise UnsatError(f""" Failed to synthesize the expression "{goal.cpp_type()} {goal.name}". When I failed, the following bindings were available in the context: {ctx_desc} This probably means there is a missing rule in the rules of torchgen.api.translate. Check this module for more information. """) # A shitty backtracking search implementation. It's shitty because it # does backtracking via stack (bad idea!) and for the most part tries to # avoid backtracking. In particular, if # direct=True, we won't try to do any fancy synthesis, just trivial # conversions (e.g., "T a" is OK for "const T& a"). So all of the # existing rules in this function simply try to solve immediately, # and bail if things don't work out. def solve(goal: NamedCType, *, direct: bool) -> str: def direct_solve(goal: NamedCType) -> str: return solve(goal, direct=True) if goal in ctx: # Trivial return ctx[goal] # const & is satisfied with mutable & if isinstance(goal.type, ConstRefCType): try: # WARNING: not strictly decreasing; be careful not # to add a direct conversion that goes satisfies # mutable& with const& return solve(NamedCType(goal.name, MutRefCType(goal.type.elem)), direct=direct) except UnsatError: pass # mutable & is satisfied with value if isinstance(goal.type, MutRefCType): try: return solve(NamedCType(goal.name, goal.type.elem), direct=direct) except UnsatError: pass if direct: unsat(goal) # For now, all of these rules are mutually exclusive. if goal == NamedCType("memory_format", OptionalCType(BaseCType(memoryFormatT))): memory_format = direct_solve( NamedCType( SpecialArgName.possibly_redundant_memory_format, OptionalCType(BaseCType(memoryFormatT)), )) # No need to join "memory_format" and "options" if the target API takes "options" directly. # Otherwise it will cause the redundant memory_format error. if options_ctype in goal_ctypes: return memory_format try: options = direct_solve(options_ctype) return f"c10::impl::check_tensor_options_and_extract_memory_format({options}, {memory_format})" except UnsatError: return memory_format elif goal == NamedCType("options", BaseCType(tensorOptionsT)): dtype = direct_solve( NamedCType("dtype", OptionalCType(BaseCType(scalarTypeT)))) pin_memory = direct_solve( NamedCType("pin_memory", OptionalCType(BaseCType(boolT)))) device = direct_solve( NamedCType("device", OptionalCType(BaseCType(deviceT)))) layout = direct_solve( NamedCType("layout", OptionalCType(BaseCType(layoutT)))) return f"TensorOptions().dtype({dtype}).layout({layout}).device({device}).pinned_memory({pin_memory})" elif goal == NamedCType("dtype", OptionalCType(BaseCType(scalarTypeT))): try: options = direct_solve(options_ctype) return f"optTypeMetaToScalarType({options}.dtype_opt())" except UnsatError: out_tensor = direct_solve(out_tensor_ctype) return f"{out_tensor}.scalar_type()" elif goal == NamedCType("layout", OptionalCType(BaseCType(layoutT))): try: options = direct_solve(options_ctype) return f"{options}.layout_opt()" except UnsatError: out_tensor = direct_solve(out_tensor_ctype) return f"{out_tensor}.layout()" elif goal == NamedCType("device", OptionalCType(BaseCType(deviceT))): try: options = direct_solve(options_ctype) return f"{options}.device_opt()" except UnsatError: out_tensor = direct_solve(out_tensor_ctype) return f"{out_tensor}.device()" elif goal == NamedCType("pin_memory", OptionalCType(BaseCType(boolT))): try: options = direct_solve(options_ctype) return f"{options}.pinned_memory_opt()" except UnsatError: # If we're calling a factory op from its out= variant, # We don't actually care about the value of pin_memory. out_tensor = direct_solve(out_tensor_ctype) return "c10::nullopt" # We can always do translations from value types to reference types, like vector<int> -> IntArrayRef elif goal.type == BaseCType(intArrayRefT): try: return direct_solve(NamedCType(goal.name, longVec_ctype)) except UnsatError: # We can also go SymIntArrayRef -> IntArrayRef symIntArrayRef_type = direct_solve( NamedCType(goal.name, BaseCType(symIntArrayRefT))) return f"c10::asIntArrayRefSlow({symIntArrayRef_type})" elif goal.type == BaseCType(symIntArrayRefT): return direct_solve(NamedCType(goal.name, longSymVec_ctype)) elif goal.type == BaseCType(longT): symInt_type = direct_solve( NamedCType(goal.name, BaseCType(SymIntT))) return f"{symInt_type}.expectInt()" elif goal.type == BaseCType(optionalIntArrayRefT): return direct_solve(NamedCType(goal.name, optionalLongVec_ctype)) elif goal.type == BaseCType(optionalScalarRefT): return direct_solve(NamedCType(goal.name, optionalScalar_ctype)) elif goal.type == BaseCType(optionalTensorRefT): return direct_solve(NamedCType(goal.name, optionalTensor_ctype)) # Note [translation from C++ reference to value types] # The below cases are all for when we have an argument with a reference type, # and a corresponding goal with a value type. # These are needed when we populate the inputs to a lambda capture and we need # to guarantee the lifetime of each captured argument. # We guard it with an explicit kwarg because converting to a value type is expensive # (O(n)) to convert from IntArrayRef to vector<int>), # so the caller of translate() should be explicit that they need it. if allow_expensive_conversions: if goal.type == VectorCType(BaseCType(longT)): intArrayRef_ctype = NamedCType(goal.name, BaseCType(intArrayRefT)) argname = direct_solve(intArrayRef_ctype) return f"{argname}.vec()" if goal.type == VectorCType(BaseCType(SymIntT)): symIntArrayRef_ctype = NamedCType(goal.name, BaseCType(symIntArrayRefT)) argname = direct_solve(symIntArrayRef_ctype) return f"{argname}.vec()" elif goal.type == OptionalCType(VectorCType(BaseCType(longT))): optionalIntArrayRef_ctype = NamedCType( goal.name, BaseCType(optionalIntArrayRefT)) argname = direct_solve(optionalIntArrayRef_ctype) return f"{argname}.has_value() ? c10::make_optional({argname}->vec()) : c10::nullopt" elif goal.type == OptionalCType(BaseCType(scalarT)): optionalScalarRef_ctype = NamedCType( goal.name, BaseCType(optionalScalarRefT)) argname = direct_solve(optionalScalarRef_ctype) return f"{argname}.has_value() ? c10::make_optional({argname}) : c10::nullopt" elif goal.type == OptionalCType(BaseCType(scalarT)): optionalTensorRef_ctype = NamedCType( goal.name, BaseCType(optionalTensorRefT)) argname = direct_solve(optionalTensorRef_ctype) return f"{argname}.has_value() ? c10::make_optional({argname}) : c10::nullopt" # Technically, we also need to handle cases of C++ containers holding reference types. # But there currently aren't any ops that require lambda capture codegen # With arguments like std::vector<IntArrayRef>. # If that changes, we'll have to add the translation here. # We allow const casting on tensors, since const-correctness is a bit broken for at::Tensor. # We could probably generalize this to non-tensor types too. if goal.type == MutRefCType(BaseCType(tensorT)): const_ref_tensor_ctype = NamedCType( goal.name, ConstRefCType(BaseCType(tensorT))) argname = direct_solve(const_ref_tensor_ctype) return f"const_cast<Tensor&>({argname})" unsat(goal) return [Expr(solve(g, direct=False), g) for g in goal_ctypes]
def gen_one(self, f: NativeFunction) -> Optional[str]: assert not f.manual_kernel_registration if (self.target is Target.REGISTRATION and not self.selector.is_native_function_selected(f)): return None # TODO: Now, there is something interesting going on here. In the code below, # we generate CompositeExplicitAutograd implementations of functional and inplace # based on the out implementation. But in fact, out is definable by # functional too (just not very efficiently), and this is honestly the # MORE likely situation for a backend implementor. How do we pick? # Well, taking a page from Haskell type classes and default methods, # we could conceivably register a circular definition (out in terms # of functional, and functional in terms of out) and just require # someone to implement one or the other. We'd have to do a little bit # of work to not register one of these "weak" definitions unless there # is a strong definition somewhere in the DAG! So it's not implemented yet. if (self.backend_index.dispatch_key == DispatchKey.CompositeExplicitAutograd and f.func.kind() is SchemaKind.out): # Never generate a default implementation for out, that's what you # have to define as a backend implementor return None # Note [Direct dispatch bindings] # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ # Signature of the non-dispatched function we'll expose in a header # (e.g., at::cpu::add). We don't generate methods (TODO: do this # when CPUTensor class is a thing); nor do we generate fallback # bindings for manual_cpp_binding functions. cpp_sig_group = CppSignatureGroup.from_native_function( f, method=False, fallback_binding=False) # Signature of the wrapper function we'll register to the dispatcher sig = NativeSignature(f.func, prefix="wrapper_") if self.target is Target.NAMESPACED_DECLARATION: result = f"TORCH_API {cpp_sig_group.signature.decl()};\n" if cpp_sig_group.faithful_signature is not None: result += f"TORCH_API {cpp_sig_group.faithful_signature.decl()};\n" return result elif self.target is Target.NAMESPACED_DEFINITION: def generate_defn(cpp_sig: CppSignature) -> str: return f""" {cpp_sig.defn()} {{ return {sig.name()}({', '.join(e.expr for e in translate(cpp_sig.arguments(), sig.arguments()))}); }} """ result = generate_defn(cpp_sig_group.signature) if cpp_sig_group.faithful_signature is not None: result += generate_defn(cpp_sig_group.faithful_signature) return result elif self.target is Target.ANONYMOUS_DEFINITION: k = f.func.kind() # Construct the body of the wrapper function with signature sig sig_body = [] # We'll use context to keep track of any variables we've brought # into scope while generating code context: List[Union[Binding, Expr]] = list(sig.arguments()) # Initialize the class corresponding to this structured # operator; feeding it the output argument(s) if it is known if self.backend_index.dispatch_key is DispatchKey.Meta: class_name = f"structured_{meta.name(self.g)}_meta_{k.name}" parent_class = f"at::meta::structured_{meta.name(self.g)}" elif (self.backend_index.dispatch_key is DispatchKey.CompositeExplicitAutograd): # TODO: dedup this branch class_name = f"structured_{meta.name(self.g)}_default_backend_{k.name}" parent_class = f"at::meta::structured_{meta.name(self.g)}" else: metadata = self.backend_index.get_kernel(self.g) assert metadata is not None class_name = f"structured_{metadata.kernel}_{k.name}" parent_class = f"{self.cpp_namespace}::structured_{metadata.kernel}" if self.backend_index.device_guard: device_check_args = itertools.chain( f.func.arguments.out, f.func.arguments.flat_positional) sig_body.append( RegisterDispatchKey.gen_device_check( f.device_check, list(device_check_args), sig.name())) if k is SchemaKind.functional: sig_body.append(f"{class_name} op;") elif k is SchemaKind.inplace: sig_body.append(f"{class_name} op(self);") elif k is SchemaKind.out: out_args_str = ", ".join(a.name for a in f.func.arguments.out) sig_body.append(f"{class_name} op({out_args_str});") # Translate the input native arguments into structured # arguments for the meta call meta_exprs = ", ".join(e.expr for e in translate( context, structured.meta_arguments(self.g), method=False)) if self.g.out.precomputed: # If this function group has precomputed elements, the meta function # returns a struct containing them which must be saved so that it # can be unpacked when generating code to call the impl. sig_body.append(f"auto precompute = op.meta({meta_exprs});") # Put all of the contents of the precompute struct into the context # so that translate will be able to return the correct args for the # call to the impl. precomputed_values = [ *self.g.out.precomputed.replace.values(), self.g.out.precomputed.add, ] for precomputed_elems in precomputed_values: for arg in precomputed_elems: context.append( Expr( expr=f"precompute.{arg.name}", type=structured.argument_type(arg, binds=arg.name), )) # Add a use of the precompute struct so FB internal compilers don't # complain that there is an unused variable. sig_body.append("(void)precompute;") else: sig_body.append(f"op.meta({meta_exprs});") # After running meta, op.outputs_ is guaranteed to be valid; # add it to the context out_args = structured.out_arguments(self.g) maybe_star = "*" if k is SchemaKind.functional else "" for i, out_arg in enumerate(out_args): assert ConstRefCType(BaseCType(tensorT)) == out_arg.nctype.type context.append( Expr( expr=f"{maybe_star}op.outputs_[{i}]", # TODO: Stop hardcoding that the output type is a Tensor. Note # that for the codegen here this is fine because outputs_ is # hardcoded to be tensor already type=NamedCType(out_arg.nctype.name, MutRefCType(BaseCType(tensorT))), )) # With the expanded context, do the impl call (if not a meta # function) if self.backend_index.dispatch_key == DispatchKey.CompositeExplicitAutograd: # TODO: https://github.com/pytorch/pytorch/issues/53023 out_sig_group = CppSignatureGroup.from_native_function( self.g.out, method=False, fallback_binding=f.manual_cpp_binding) out_sig = out_sig_group.most_faithful_signature() api_name = out_sig.name() out_exprs = ", ".join(e.expr for e in translate( context, out_sig.arguments(), method=False)) # TODO: I think this means structured won't work with method # only functions (but maybe you're saved by faithful? iunno.) # NB: Originally I wrote this as an at::redispatch call, but # I got in trouble because that meant I needed a DispatchKeySet # in the wrapper function, which meant I needed a DispatchKeySet # in the DispatchKeyFunctions declarations, but the defined API # there does NOT permit a dispatch key set. I think you can # probably unwind this by calling some function to do the TLS # fetch and get the DispatchKeySet when you don't have it, but # I didn't do it for this version sig_body.append(f"at::{api_name}({out_exprs});") elif self.backend_index.dispatch_key != DispatchKey.Meta: impl_exprs = ", ".join(e.expr for e in translate( context, structured.impl_arguments(self.g), method=False)) sig_body.append(f"op.impl({impl_exprs});") # Destructively return the final tensors # TODO: Do this in translate instead if k is SchemaKind.functional: if len(f.func.returns) == 1: ret_expr = "std::move(op.outputs_[0]).take()" # small optimization else: moved = ", ".join(f"std::move(op.outputs_[{i}]).take()" for i in range(len(f.func.returns))) ret_expr = f"std::make_tuple({moved})" elif k is SchemaKind.inplace: ret_expr = "self" elif k is SchemaKind.out: if len(f.func.returns) == 1: ret_expr = f.func.arguments.out[0].name else: refs = ", ".join(a.name for a in f.func.arguments.out) ret_expr = f"std::forward_as_tuple({refs})" sig_body.append(f"return {ret_expr};") sig_body_str = "\n".join(sig_body) # For an overview of what this template code looks like, see # https://github.com/pytorch/rfcs/pull/9 return f"""\ {self.gen_class( f, k, class_name=class_name, parent_class=parent_class, generate_super=self.g.out.structured_inherits is not None )} {sig.defn()} {{ {sig_body_str} }} """ elif self.target is Target.REGISTRATION: return f'm.impl("{f.func.name}", TORCH_FN({sig.name()}));' else: assert_never(self.target) # Silence mypy's "Missing return statement" error return None
def compute_ufunc_cpu_dtype_body( g: NativeFunctionsGroup, dtype: ScalarType, inner_loops: Dict[UfuncKey, UfuncSignature], parent_ctx: Sequence[Binding], ) -> str: assert UfuncKey.CPUScalar in inner_loops, f"{dtype}, {inner_loops.keys()}" assert inner_loops.keys() <= {UfuncKey.CPUScalar, UfuncKey.CPUVector} scalar_loop = inner_loops[UfuncKey.CPUScalar] vec_loop = None if UfuncKey.CPUVector in inner_loops: vec_loop = inner_loops[UfuncKey.CPUVector] # NB: We DON'T use translate here, because translate is # incapable of CSE'ing the scalar accesses in case it is also # used by Vectorized; also, the unpacking here is very simple # and only affects Scalar; everything else is implicitly captured # by the lambda # Setup scalar in scope body = [] ctx = [] for b in parent_ctx: if isinstance(b.argument, Argument) and b.argument.type != BaseType(BaseTy.Scalar): continue body.append(f"auto _s_{b.name} = {b.name}.to<scalar_t>();") ctx.append( Expr(f"_s_{b.name}", NamedCType(b.nctype.name, BaseCType(scalar_t)))) if vec_loop is not None: for b in parent_ctx: if isinstance( b.argument, Argument) and b.argument.type != BaseType(BaseTy.Scalar): continue body.append( f"auto _v_{b.name} = at::vec::Vectorized<scalar_t>(_s_{b.name});" ) ctx.append( Expr( f"_v_{b.name}", NamedCType(b.nctype.name, VectorizedCType(BaseCType(scalar_t))), )) # Setup lambda signature # NB: simplified version of ufunctor_arguments scalar_bindings = [] vec_bindings = [] for a in g.functional.func.arguments.flat_non_out: if not a.type.is_tensor_like(): continue assert a.type == BaseType(BaseTy.Tensor) scalar_bindings.append( Binding( name=a.name, nctype=NamedCType(a.name, BaseCType(scalar_t)), argument=a, )) if vec_loop is not None: vec_bindings.append( Binding( name=a.name, nctype=NamedCType(a.name, VectorizedCType(BaseCType(scalar_t))), argument=a, )) def with_ctx(b: Sequence[Binding]) -> List[Union[Expr, Binding]]: r: List[Union[Expr, Binding]] = [] r.extend(ctx) r.extend(b) return r body_str = "\n".join(body) if vec_loop is not None: return f""" {body_str} cpu_kernel_vec(iter, [=]({', '.join(b.decl() for b in scalar_bindings)}) {{ return {scalar_loop.call(with_ctx(scalar_bindings))}; }}, [=]({', '.join(b.decl() for b in vec_bindings)}) {{ return {vec_loop.call(with_ctx(vec_bindings))}; }} ); """ else: return f"""