From 1304172a9364ebff64f39604ef00fa0c5b5cbdfb Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Tue, 19 May 2026 08:53:13 -0400
Subject: [PATCH] Modified repack

---
 Grid/lattice/Lattice_reduction_gpu.h | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/Grid/lattice/Lattice_reduction_gpu.h b/Grid/lattice/Lattice_reduction_gpu.h
index 5cc8636f..7fa80bcd 100644
--- a/Grid/lattice/Lattice_reduction_gpu.h
+++ b/Grid/lattice/Lattice_reduction_gpu.h
@@ -261,11 +261,9 @@ inline void sumD_gpu_reduce_words(const vobj *lat, Integer osites,
 #ifdef GRID_REDUCTION_TIMING
   RealD t_pack = -usecond();
 #endif
-  accelerator_for(ss, osites, 1, {
-    Bundle b;
-    for (int k = 0; k < R; k++)
-      b._internal[k] = idat[ss * words + base + k];
-    buf_p[ss] = b;
+  constexpr int Nsimd = vobj::Nsimd();
+  accelerator_for2d(k, R, ss, osites, Nsimd, {
+    coalescedWrite(buf_p[ss]._internal[k], coalescedRead(idat[ss * words + base + k]));
   });
 #ifdef GRID_REDUCTION_TIMING
   t_pack += usecond();