Merge branch 'ExtendIndexedDomainKernel' into 'master'

Allow index expression for indexed domain kernel

See merge request pycodegen/pystencils!416

src/pystencils/gpu/kernelcreation.py

@@ -72,9 +72,11 @@ def create_cuda_kernel(assignments: NodeCollection, config: CreateKernelConfig):
 
     if len(indexed_elements) > 0:
         common_indexed_element = get_common_indexed_element(indexed_elements)
+        index = common_indexed_element.indices[0].atoms(TypedSymbol)
+        assert len(index) == 1, "index expressions must only contain one symbol representing the index"
         indexing = indexing_creator(iteration_space=(slice(0, common_indexed_element.shape[0], 1), *iteration_space),
                                     data_layout=common_field.layout)
-        extended_ctrs = [common_indexed_element.indices[0], *loop_counter_symbols]
+        extended_ctrs = [index.pop(), *loop_counter_symbols]
         loop_counter_assignments = indexing.get_loop_ctr_assignments(extended_ctrs)
     else:
         indexing = indexing_creator(iteration_space=iteration_space, data_layout=common_field.layout)

src/pystencils/transformations.py

@@ -276,8 +276,10 @@ def add_outer_loop_over_indexed_elements(loop_node: ast.Block) -> ast.Block:
     if len(indexed_elements) == 0:
         return loop_node
     reference_element = get_common_indexed_element(indexed_elements)
+    index = reference_element.indices[0].atoms(TypedSymbol)
+    assert len(index) == 1, "index expressions must only contain one symbol representing the index"
     new_loop = ast.LoopOverCoordinate(loop_node, 0, 0,
-                                      reference_element.shape[0], 1, custom_loop_ctr=reference_element.indices[0])
+                                      reference_element.shape[0], 1, custom_loop_ctr=index.pop())
     return ast.Block([new_loop])
 
 

tests/test_indexed_kernels.py

@@ -64,15 +64,15 @@ def test_indexed_domain_kernel(index_size, array_size, target, dtype):
     src = sp.IndexedBase(TypedSymbol(f"_data_{f.name}", dtype=const_pointer_type), shape=index_src)
     dst = sp.IndexedBase(TypedSymbol(f"_data_{g.name}", dtype=pointer_type), shape=index_dst)
 
-    update_rule = [ps.Assignment(FieldPointerSymbol("f", dtype, const=True), src[index]),
-                   ps.Assignment(FieldPointerSymbol("g", dtype, const=False), dst[index]),
+    update_rule = [ps.Assignment(FieldPointerSymbol("f", dtype, const=True), src[index + 1]),
+                   ps.Assignment(FieldPointerSymbol("g", dtype, const=False), dst[index + 1]),
                    ps.Assignment(g.center, f.center)]
 
     ast = ps.create_kernel(update_rule, target=target)
 
     code = ps.get_code_str(ast)
-    assert f"const {dtype.c_name} * RESTRICT _data_f = (({dtype.c_name} * RESTRICT const)(_data_f[index]));" in code
-    assert f"{dtype.c_name} * RESTRICT  _data_g = (({dtype.c_name} * RESTRICT )(_data_g[index]));" in code
+    assert f"const {dtype.c_name} * RESTRICT _data_f = (({dtype.c_name} * RESTRICT const)(_data_f[index + 1]));" in code
+    assert f"{dtype.c_name} * RESTRICT  _data_g = (({dtype.c_name} * RESTRICT )(_data_g[index + 1]));" in code
 
     if target == Target.CPU:
         assert code.count("for") == f.spatial_dimensions + 1