From: Szilárd Páll <pall.szilard@gmail.com>
Date: Mon, 2 Dec 2019 18:29:28 +0000 (+0100)
Subject: Allow overlapping CPU force H2D with compute
X-Git-Url: http://biod.pnpi.spb.ru/gitweb/?a=commitdiff_plain;h=86a27bc2bba63500b679b30c8ad7913fc5425680;p=alexxy%2Fgromacs.git

Allow overlapping CPU force H2D with compute

The reduction orchestration code already uses explicit sync event
in all cases and StateGpu implements the ability to schedule force
H2D in a separate stream for the "All" locality.
Hence, this change switches for non-DD runs the CPU force H2D to be done
in the update stream to allow overlap with force work in the local
stream.

Refs #3170 #3029

Change-Id: Iceb9aac395335c062109d552d3f0289688a9c75f
---

diff --git a/src/gromacs/mdlib/sim_util.cpp b/src/gromacs/mdlib/sim_util.cpp
index 832ff21620..86549f915b 100644
--- a/src/gromacs/mdlib/sim_util.cpp
+++ b/src/gromacs/mdlib/sim_util.cpp
@@ -1706,9 +1706,16 @@ void do_force(FILE*                               fplog,
             //   These should be unified.
             if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
             {
-                stateGpu->copyForcesToGpu(forceWithShift, AtomLocality::Local);
+                // Note: AtomLocality::All is used for the non-DD case because, as in this
+                // case copyForcesToGpu() uses a separate stream, it allows overlap of
+                // CPU force H2D with GPU force tasks on all streams including those in the
+                // local stream which would otherwise be implicit dependencies for the
+                // transfer and would not overlap.
+                auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
+
+                stateGpu->copyForcesToGpu(forceWithShift, locality);
                 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(
-                        AtomLocality::Local, useGpuFBufOps == BufferOpsUseGpu::True));
+                        locality, useGpuFBufOps == BufferOpsUseGpu::True));
             }
             if (useGpuForcesHaloExchange)
             {