Relax OpenCL gather kernel barrier on AMD

Not needed on arch with >32 execution width.

Refs #2519

Change-Id: I4c8f2cff405cd3044bd60b99f01bcdd918dc5d0e

diff --git a/src/gromacs/ewald/pme-gather.clh b/src/gromacs/ewald/pme-gather.clh
index e795d4dd7086f7862b46327cd04a61481a17cb1f..9061701fcc73a203290e85809d543d346aad69ec 100644 (file)
--- a/src/gromacs/ewald/pme-gather.clh
+++ b/src/gromacs/ewald/pme-gather.clh
@@ -1,7 +1,7 @@
 /*
  * This file is part of the GROMACS molecular simulation package.
  *
- * Copyright (c) 2018, by the GROMACS development team, led by
+ * Copyright (c) 2018,2019, by the GROMACS development team, led by
  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
  * and including many others, as listed in the AUTHORS file in the
  * top-level source directory and at http://www.gromacs.org.
@@ -122,11 +122,15 @@ inline void reduce_atom_forces(__local float * __restrict__  sm_forces,
         int elementIndex = smemReserved + lineIndex;
         // Store input force contributions
         sm_forceReduction[elementIndex] = (dimIndex == XX) ? fx : (dimIndex == YY) ? fy : fz;
-        /* This barrier was not needed in CUDA. Different OpenCL compilers might have different ideas
+
+#if !defined(_AMD_SOURCE_)
+        /* This barrier was not needed in CUDA, nor is it needed on AMD GPUs.
+         * Different OpenCL compilers might have different ideas
          * about #pragma unroll, though. OpenCL 2 has _attribute__((opencl_unroll_hint)).
          * #2519
          */
         barrier(CLK_LOCAL_MEM_FENCE);
+#endif
 
         // Reduce to fit into smemPerDim (warp size)
 #pragma unroll