From 2da09ae99b0ede9f674d8f6f70732fe154882f22 Mon Sep 17 00:00:00 2001
From: david clarke <clarke.davida@gmail.com>
Date: Tue, 6 Feb 2024 18:40:13 -0700
Subject: [PATCH] acceleration compiles and doesn't break scalar mode

---
 Grid/qcd/smearing/HISQSmearing.h | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/Grid/qcd/smearing/HISQSmearing.h b/Grid/qcd/smearing/HISQSmearing.h
index 0deb080d..2ae11bbb 100644
--- a/Grid/qcd/smearing/HISQSmearing.h
+++ b/Grid/qcd/smearing/HISQSmearing.h
@@ -153,18 +153,18 @@ public:
         // This loop handles 3-, 5-, and 7-link constructs, minus Lepage and Naik.
         for(int mu=0;mu<Nd;mu++) {
 
-            // Create the accessors
-            autoView(U_v       , Ughost       , AcceleratorRead);
-            autoView(U_fat_v   , Ughost_fat   , CpuWrite);
-            autoView(U_3link_v , Ughost_3link , CpuWrite);
-            autoView(U_5linkA_v, Ughost_5linkA, CpuWrite);
-            autoView(U_5linkB_v, Ughost_5linkB, CpuWrite);
-
             // TODO: This approach is slightly memory inefficient. It uses 25% extra memory 
             Ughost_3link =Zero();
             Ughost_5linkA=Zero();
             Ughost_5linkB=Zero();
 
+            // Create the accessors
+            autoView(U_v       , Ughost       , AcceleratorRead);
+            autoView(U_fat_v   , Ughost_fat   , AcceleratorWrite);
+            autoView(U_3link_v , Ughost_3link , AcceleratorWrite);
+            autoView(U_5linkA_v, Ughost_5linkA, AcceleratorWrite);
+            autoView(U_5linkB_v, Ughost_5linkB, AcceleratorWrite);
+
             // We infer some types that will be needed in the calculation.
             typedef decltype(gStencil.GetEntry(0,0)) stencilElement;
             typedef decltype(coalescedReadGeneralPermute(U_v[0](0),gStencil.GetEntry(0,0)->_permute,Nd)) U3matrix;