Merge remote-tracking branch 'upstream/develop' into omp

include/kernels/gauge_stout.cuh

@@ -151,13 +151,8 @@ namespace quda
       dir = dir + (dir >= arg.dir_ignore);
 
       Link U, Q;
-#ifdef QUDA_OMPTARGET_THREAD_ARRAY_SIMPLE
-      Link Stap;
-      Link Rect;
-#else
-      ThreadLocalCache<Link, 0, computeStapleRectangleOps> Stap;
-      ThreadLocalCache<Link, 0, decltype(Stap)> Rect; // offset by Stap type to ensure non-overlapping allocations
-#endif
+      typename OvrImpSTOUTOps<Arg>::StapCacheT Stap {*this};
+      typename OvrImpSTOUTOps<Arg>::RectCacheT Rect {*this};
 
       // This function gets stap = S_{mu,nu} i.e., the staple of length 3,
       // and the 1x2 and 2x1 rectangles of length 5. From the following paper:

include/kernels/gauge_utils.cuh

@@ -1,7 +1,7 @@
 #include <gauge_field_order.h>
 #include <index_helper.cuh>
 #include <quda_matrix.h>
-#include <shared_memory_cache_helper.h>
+#include <thread_local_cache.h>
 #include <thread_array.h>
 
 namespace quda

include/kernels/gauge_wilson_flow.cuh

@@ -83,14 +83,9 @@ namespace quda
       // This function gets stap = S_{mu,nu} i.e., the staple of length 3,
       // and the 1x2 and 2x1 rectangles of length 5. From the following paper:
       // https://arxiv.org/abs/0801.1165
-#ifdef QUDA_OMPTARGET_THREAD_ARRAY_SIMPLE
-      Link Stap;
-      Link Rect;
-#else
-      ThreadLocalCache<Link, 0, computeStapleRectangleOps> Stap;
-      ThreadLocalCache<Link, 0, decltype(Stap)> Rect; // offset by Stap type to ensure non-overlapping allocations
-#endif
-      computeStapleRectangle(arg, x, arg.E, parity, dir, Stap, Rect, Arg::wflow_dim);
+      typename computeStapleOpsWF<Arg>::StapOp Stap {ftor};
+      typename computeStapleOpsWF<Arg>::RectOp Rect {ftor};
+      computeStapleRectangle(ftor, x, arg.E, parity, dir, Stap, Rect, Arg::wflow_dim);
       Z = arg.coeff1x1 * static_cast<const Link &>(Stap) + arg.coeff2x1 * static_cast<const Link &>(Rect);
     }
     return Z;