Fitch local fetch op in CFEM quadrature.

examples/dg.py

-# FIXME NOT UPDATED YET FOR NEW-STYLE LOOPY!
+raise NotImplementedError("NOT UPDATED YET FOR NEW-STYLE LOOPY!") # FIXME
 
 
 

examples/quadrature.py

@@ -60,7 +60,8 @@ def build_mass_mat_maker(ctx_factory=cl.create_some_context):
             outer_slab_increments=(0,0))
 
     # fix reg prefetch
-    knl = lp.add_prefetch(knl, "det_j", ["c_inner"])
+    knl = lp.add_prefetch(knl, "det_j", ["c_inner"],
+            loc_fetch_axes={0: (0, 1)})
 
     #ilp = 4
     #knl = lp.split_dimension(knl, "i", 2, outer_tag="g.0", inner_tag="l.1")

loopy/__init__.py

@@ -123,7 +123,7 @@ def add_prefetch(kernel, input_access_descr, fetch_dims, loc_fetch_axes={}):
             input_vector=ivec,
             index_expr=iexpr,
             fetch_dims=fetch_dims,
-            loc_fetch_axes={})
+            loc_fetch_axes=loc_fetch_axes)
 
     return kernel.copy(prefetch=new_prefetch)
 

loopy/codegen/prefetch.py

@@ -303,11 +303,17 @@ def generate_prefetch_code(kernel, sched_index, exec_domain):
 
     from loopy.kernel import TAG_WORK_ITEM_IDX
     knl_work_item_inames = kernel.ordered_inames_by_tag_type(TAG_WORK_ITEM_IDX)
+    used_kernel_work_item_inames = []
 
     for realization_dim_idx, loc_fetch_axis_list in \
-            getattr(pf, "loc_fetch_axes", {}).iteritems():
-        realization_inames[realization_dim_idx] = [knl_work_item_inames.pop(axis)
+            pf.loc_fetch_axes.iteritems():
+        loc_fetch_inames = [knl_work_item_inames[axis]
             for axis in loc_fetch_axis_list]
+        realization_inames[realization_dim_idx] = loc_fetch_inames
+        used_kernel_work_item_inames.extend(loc_fetch_inames)
+
+    for inm in used_kernel_work_item_inames:
+        knl_work_item_inames.remove(inm)
 
     # }}}