From 33cdb10598caa90acaadb568d4e869174ad70ec4 Mon Sep 17 00:00:00 2001
From: rhaas <rhaas@b61c5cb5-eaca-4651-9a7a-d64986f99364>
Date: Fri, 2 Aug 2013 21:34:14 +0000
Subject: add code to align allocated memory for vectorization

git-svn-id: http://svn.cactuscode.org/arrangements/CactusPUGH/PUGH/trunk@524 b61c5cb5-eaca-4651-9a7a-d64986f99364
---
 configuration.ccl   |  4 ++++
 interface.ccl       |  2 ++
 src/Storage.c       | 14 +++++++++++---
 src/include/pughi.h |  2 ++
 src/make.code.defn  |  2 +-
 5 files changed, 20 insertions(+), 4 deletions(-)

diff --git a/configuration.ccl b/configuration.ccl
index 1a5e567..db93a45 100644
--- a/configuration.ccl
+++ b/configuration.ccl
@@ -10,3 +10,7 @@ PROVIDES PUGH
 OPTIONAL MPI
 {
 }
+
+OPTIONAL_IFACTIVE Vectors
+{
+}
diff --git a/interface.ccl b/interface.ccl
index 95c3b2a..734ec17 100644
--- a/interface.ccl
+++ b/interface.ccl
@@ -6,3 +6,5 @@ implements: Driver
 INCLUDE HEADER: Register.h in pugh_Register.h
 
 USES INCLUDE: pugh_Register.h
+
+USES INCLUDE: Vectors.h
diff --git a/src/Storage.c b/src/Storage.c
index cbdf47d..0146b07 100644
--- a/src/Storage.c
+++ b/src/Storage.c
@@ -628,9 +628,17 @@ static int PUGH_EnableGArrayGroupStorage (pGH *pughGH,
       else if (! my_padding_active)
       {
         /* Easy case. */
-        GA->data = malloc ((size_t) GA->extras->npoints * GA->varsize
-			   * GA->vector_size);
-        GA->padddata = GA->data;
+        size_t align = PUGH_GetVectorSize() * GA->varsize;
+        GA->padddata = malloc ((size_t) GA->extras->npoints * GA->varsize
+			      * GA->vector_size + align-1);
+        if (((size_t) GA->padddata) % align)
+        {
+          GA->data = (char *) GA->padddata + align - (((size_t) GA->padddata) % align);
+        }
+        else
+        {
+          GA->data = GA->padddata;
+        }
       }
       else
       {
diff --git a/src/include/pughi.h b/src/include/pughi.h
index b934690..67ec409 100644
--- a/src/include/pughi.h
+++ b/src/include/pughi.h
@@ -106,6 +106,8 @@ int PUGH_SyncGArray(pGA *GA);
 
 int PUGH_NumTimeLevels(const pGH *pughGH, int var);
 
+size_t PUGH_GetVectorSize(void);
+
 #ifdef CCTK_MPI
 void PostSendGA     (pGH *pughGH, int dir, pComm *comm);
 void PostReceiveGA  (pGH *pughGH, int dir, pComm *comm);
diff --git a/src/make.code.defn b/src/make.code.defn
index 4779272..bcb4e70 100644
--- a/src/make.code.defn
+++ b/src/make.code.defn
@@ -4,4 +4,4 @@
 SRCS =  Startup.c GHExtension.c Overloadables.c PughUtils.c LoadAware.c       \
         Evolve.c Comm.c Storage.c SetupPGH.c SetupGroup.c SetupPGV.c          \
         PostSendGA.c PostReceiveGA.c FinishReceiveGA.c Registration.c         \
-        P2LMappings.c Topology.c
+        P2LMappings.c Topology.c Vectors.cc
-- 
cgit v1.2.3