generalize parts of vector dtypes implementation

loopy/codegen/instruction.py

@@ -127,6 +127,13 @@ def generate_assignment_instruction_code(codegen_state, insn):
             raise UnvectorizableError(
                     "LHS is scalar, RHS is vector, cannot assign")
 
+        if (lhs_is_vector
+                and (not rhs_is_vector)
+                and (not
+                     kernel.target.broadcasts_scalar_assignment_to_vec_types)):
+            raise UnvectorizableError(
+                    "LHS is vector, RHS is not vector, cannot assign")
+
         is_vector = lhs_is_vector
 
         del lhs_is_vector

loopy/kernel/array.py

@@ -1401,7 +1401,9 @@ def eval_expr_assert_integer_constant(i, expr):
                 # We'll do absolutely nothing here, which will result
                 # in the vector being returned.
                 pass
-
+            elif (vectorization_info is None
+                    and target.allows_non_constant_indexing_for_vec_types):
+                vector_index = eval_expr(idx)
             else:
                 idx = eval_expr_assert_integer_constant(i, idx)
 

loopy/target/c/__init__.py

@@ -1012,6 +1012,11 @@ def ast_block_class(self):
         from cgen import Block
         return Block
 
+    @property
+    def ast_comment_class(self):
+        from cgen import Comment
+        return Comment
+
     @property
     def ast_block_scope_class(self):
         return ScopingBlock
@@ -1255,6 +1260,10 @@ def emit_comment(self, s):
         from cgen import Comment
         return Comment(s)
 
+    def emit_pragma(self, s):
+        from cgen import Pragma
+        return Pragma(s)
+
     @property
     def can_implement_conditionals(self):
         return True
@@ -1334,6 +1343,10 @@ def get_dtype_registry(self):
         fill_registry_with_c99_complex_types(result)
         return DTypeRegistryWrapper(result)
 
+    @property
+    def allows_non_constant_indexing_for_vec_types(self):
+        return False
+
 
 class CASTBuilder(CFamilyASTBuilder):
     def preamble_generators(self):

loopy/target/c/codegen/expression.py

@@ -217,15 +217,24 @@ def make_var(name):
         ary = self.find_array(expr)
 
         from loopy.kernel.array import get_access_info
-        from pymbolic import evaluate
+        from pymbolic import evaluate, substitute
 
         from loopy.symbolic import simplify_using_aff
         index_tuple = tuple(
                 simplify_using_aff(self.kernel, idx) for idx in expr.index_tuple)
 
-        access_info = get_access_info(self.kernel.target, ary, index_tuple,
-                lambda expr: evaluate(expr, self.codegen_state.var_subst_map),
-                self.codegen_state.vectorization_info)
+        if self.kernel.target.allows_non_constant_indexing_for_vec_types:
+            access_info = get_access_info(self.kernel.target, ary, index_tuple,
+                                          lambda expr: substitute(
+                                              expr,
+                                              self.codegen_state.var_subst_map),
+                                          self.codegen_state.vectorization_info)
+        else:
+            access_info = get_access_info(self.kernel.target, ary, index_tuple,
+                                          lambda expr: evaluate(
+                                              expr,
+                                              self.codegen_state.var_subst_map),
+                                          self.codegen_state.vectorization_info)
 
         from loopy.kernel.data import (
                 ImageArg, ArrayArg, TemporaryVariable, ConstantArg)

loopy/target/cuda.py

@@ -265,6 +265,14 @@ def vector_dtype(self, base, count):
 
     # }}}
 
+    @property
+    def allows_non_constant_indexing_for_vec_types(self):
+        return False
+
+    @property
+    def broadcasts_scalar_assignment_to_vec_types(self):
+        return True
+
 # }}}
 
 

loopy/target/ispc.py

@@ -199,6 +199,10 @@ def get_dtype_registry(self):
 
     # }}}
 
+    @property
+    def allows_non_constant_indexing_for_vec_types(self):
+        return False
+
 
 class ISPCASTBuilder(CFamilyASTBuilder):
     def _arg_names_and_decls(self, codegen_state):

loopy/target/opencl.py

@@ -594,6 +594,14 @@ def vector_dtype(self, base, count):
                 vec.types[base.numpy_dtype, count],
                 target=self)
 
+    @property
+    def allows_non_constant_indexing_for_vec_types(self):
+        return False
+
+    @property
+    def broadcasts_scalar_assignment_to_vec_types(self):
+        return True
+
 # }}}
 
 

loopy/target/python.py

@@ -218,6 +218,11 @@ def ast_block_scope_class(self):
         # and delete the implementation above.
         return Collection
 
+    @property
+    def ast_comment_class(self):
+        from genpy import Comment
+        return Comment
+
     def emit_sequential_loop(self, codegen_state, iname, iname_dtype,
             lbound, ubound, inner):
         ecm = codegen_state.expression_to_code_mapper