From 56628fe2b6438327cec6b928f6b0c8fc6fc1a141 Mon Sep 17 00:00:00 2001
From: stamoor <stamoor@f3b2605a-c512-4ea7-a41b-209d697bcdaa>
Date: Fri, 26 Aug 2016 18:17:16 +0000
Subject: [PATCH] Adding Kokkos warning
git-svn-id: svn://svn.icms.temple.edu/lammps-ro/trunk@15505 f3b2605a-c512-4ea7-a41b-209d697bcdaa
---
src/finish.cpp | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/src/finish.cpp b/src/finish.cpp
index 8fff3a8a54..8670987573 100644
--- a/src/finish.cpp
+++ b/src/finish.cpp
@@ -33,6 +33,7 @@
#include "neigh_request.h"
#include "output.h"
#include "memory.h"
+#include "error.h"
#ifdef LMP_USER_OMP
#include "modify.h"
@@ -515,6 +516,13 @@ void Finish::end(int flag)
}
#endif
+ if (lmp->kokkos && lmp->kokkos->ngpu > 0)
+ if (const char* env_clb = std::getenv("CUDA_LAUNCH_BLOCKING"))
+ if (!(strcmp(env_clb,"1") == 0)) {
+ error->warning(FLERR,"Timing breakdown may not be accurate since GPU/CPU overlap is enabled. "
+ "Using 'export CUDA_LAUNCH_BLOCKING=1' will give an accurate timing breakdown but will reduce performance");
+ }
+
// FFT timing statistics
// time3d,time1d = total time during run for 3d and 1d FFTs
// loop on timing() until nsample FFTs require at least 1.0 CPU sec
--
GitLab