From 77ba9d53c2a1fa7c4fd1db7eccb1578b5b5960c4 Mon Sep 17 00:00:00 2001
From: edoapra <edoardo.apra@gmail.com>
Date: Thu, 22 Mar 2018 18:06:39 -0700
Subject: [PATCH] same as commit 76d5bddd5c3dfdef76beaab8222231624eb75e89:
 Split ga_acc  in smaller ga_acc on MPI-PR since gives large performance
 improvement on NERSC Cori

---
 src/ccsd/GNUmakefile  |  4 ++++
 src/ccsd/moints_trp.F | 16 +++++++++++++++-
 2 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/src/ccsd/GNUmakefile b/src/ccsd/GNUmakefile
index dcac6a4d85..acba2a3793 100644
--- a/src/ccsd/GNUmakefile
+++ b/src/ccsd/GNUmakefile
@@ -110,6 +110,10 @@ ifeq ($(HAVE_SET_GA_PROPERTY),Y)
 endif
 
 
+ifeq ($(ARMCI_NETWORK),MPI-PR)
+   LIB_DEFINES += -DACC_STRIPS
+endif
+
  LIBRARY = libccsd.a
 
 include ../config/makelib.h
diff --git a/src/ccsd/moints_trp.F b/src/ccsd/moints_trp.F
index 875f0fc650..3e274fff62 100644
--- a/src/ccsd/moints_trp.F
+++ b/src/ccsd/moints_trp.F
@@ -532,7 +532,14 @@ cstaggering to avoid congestion in ga-acc
 #endif
 #ifndef NOCOMMS
            ab = (a-olo)*nocc+b-ostart+1
+#ifdef ACC_STRIPS
+       do i=ilo,ihi
+          call ga_acc(g_exch,(i-1)*nvir + 1,i*nvir,ab,ab,
+     C         h1(1,i),nni,1.d0)
+       enddo
+#else           
            call ga_acc(g_exch,ivlo,ivhi,ab,ab,h1,nni,1.d0)
+#endif
 #endif
          enddo
          do b=ostart,ostart+nocc-1
@@ -581,7 +588,14 @@ cstaggering to avoid congestion in ga-acc
 #endif
 #ifndef NOCOMMS
           ab = (a-olo)*nocc+b-ostart+1
-          call ga_acc(g_exch,jvlo,jvhi,ab,ab,h2,nnj,1.d0)
+#ifdef ACC_STRIPS
+       do j=jlo,jhi
+          call ga_acc(g_exch,(j-1)*nvir + 1,j*nvir,ab,ab,
+     C         h2(1,j),nnj,1.d0)
+       enddo
+#else           
+       call ga_acc(g_exch,jvlo,jvhi,ab,ab,h2,nnj,1.d0)
+#endif
 #endif
          enddo
        enddo