Merge pull request #1295 from ggouaillardet/topic/v2.x/libnbc_fixes

v2.x: misc coll/libnbc fixes

Author: jsquyres

ompi/mca/coll/libnbc/nbc_ireduce.c

@@ -14,11 +14,12 @@
  *
  */
 
+#include "opal/include/opal/align.h"
 #include "ompi/op/op.h"
 
 #include "nbc_internal.h"
 
-static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *redbuf, int count, MPI_Datatype datatype,
+static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *redbuf, char tmpredbuf, int count, MPI_Datatype datatype,
                                       MPI_Op op, char inplace, NBC_Schedule *schedule, NBC_Handle *handle);
 static inline int red_sched_chain (int rank, int p, int root, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
                                    MPI_Op op, int ext, size_t size, NBC_Schedule *schedule, NBC_Handle *handle, int fragsize);
@@ -55,6 +56,7 @@ int ompi_coll_libnbc_ireduce(const void* sendbuf, void* recvbuf, int count, MPI_
   MPI_Aint ext;
   NBC_Schedule *schedule;
   char *redbuf=NULL, inplace;
+  char tmpredbuf = 0;
   enum { NBC_RED_BINOMIAL, NBC_RED_CHAIN } alg;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
@@ -104,8 +106,10 @@ int ompi_coll_libnbc_ireduce(const void* sendbuf, void* recvbuf, int count, MPI_
       redbuf = recvbuf;
     } else {
       /* recvbuf may not be valid on non-root nodes */
-      handle->tmpbuf = malloc (2*span);
-      redbuf = (char*) handle->tmpbuf + span - gap;
+      ptrdiff_t span_align = OPAL_ALIGN(span, datatype->super.align, ptrdiff_t);
+      handle->tmpbuf = malloc (span_align + span);
+      redbuf = (char*)span_align - gap;
+      tmpredbuf = 1;
     }
   } else {
     handle->tmpbuf = malloc (span);
@@ -142,7 +146,7 @@ int ompi_coll_libnbc_ireduce(const void* sendbuf, void* recvbuf, int count, MPI_
 
     switch(alg) {
       case NBC_RED_BINOMIAL:
-        res = red_sched_binomial(rank, p, root, sendbuf, redbuf, count, datatype, op, inplace, schedule, handle);
+        res = red_sched_binomial(rank, p, root, sendbuf, redbuf, tmpredbuf, count, datatype, op, inplace, schedule, handle);
         break;
       case NBC_RED_CHAIN:
         res = red_sched_chain(rank, p, root, sendbuf, recvbuf, count, datatype, op, ext, size, schedule, handle, segsize);
@@ -289,10 +293,10 @@ int ompi_coll_libnbc_ireduce_inter(const void* sendbuf, void* recvbuf, int count
   if (vrank == 0) rank = root; \
   if (vrank == root) rank = 0; \
 }
-static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *redbuf, int count, MPI_Datatype datatype,
+static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *redbuf, char tmpredbuf, int count, MPI_Datatype datatype,
                                       MPI_Op op, char inplace, NBC_Schedule *schedule, NBC_Handle *handle) {
   int vroot, vrank, vpeer, peer, res, maxr;
-  char *rbuf, *lbuf, *buf;
+  char *rbuf, *lbuf, *buf, tmpbuf;
   int tmprbuf, tmplbuf;
   ptrdiff_t gap;
   (void)opal_datatype_span(&datatype->super, count, &gap);
@@ -310,12 +314,12 @@ static inline int red_sched_binomial (int rank, int p, int root, const void *sen
     rbuf = (void *)(-gap);
     tmprbuf = true;
     lbuf = redbuf;
-    tmplbuf = false;
+    tmplbuf = tmpredbuf;
   } else {
     lbuf = (void *)(-gap);
     tmplbuf = true;
     rbuf = redbuf;
-    tmprbuf = false;
+    tmprbuf = tmpredbuf;
     if (inplace) {
         res = NBC_Copy(rbuf, count, datatype, ((char *)handle->tmpbuf)-gap, count, datatype, MPI_COMM_SELF);
         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -352,7 +356,7 @@ static inline int red_sched_binomial (int rank, int p, int root, const void *sen
         }
         /* swap left and right buffers */
         buf = rbuf; rbuf = lbuf ; lbuf = buf;
-        tmprbuf ^= 1; tmplbuf ^= 1;
+        tmpbuf = tmprbuf; tmprbuf = tmplbuf; tmplbuf = tmpbuf;
       }
     } else {
       /* we have to send this round */
@@ -377,9 +381,9 @@ static inline int red_sched_binomial (int rank, int p, int root, const void *sen
   /* send to root if vroot ! root */
   if (vroot != root) {
     if (0 == rank) {
-      res = NBC_Sched_send (redbuf, false, count, datatype, root, schedule, false);
+      res = NBC_Sched_send (redbuf, tmpredbuf, count, datatype, root, schedule, false);
     } else if (root == rank) {
-      res = NBC_Sched_recv (redbuf, false, count, datatype, vroot, schedule, false);
+      res = NBC_Sched_recv (redbuf, tmpredbuf, count, datatype, vroot, schedule, false);
     }
   }
 

ompi/mca/coll/libnbc/nbc_ireduce_scatter.c

@@ -16,6 +16,8 @@
  * Author(s): Torsten Hoefler <[email protected]>
  *
  */
+#include "opal/include/opal/align.h"
+
 #include "nbc_internal.h"
 
 /* an reduce_csttare schedule can not be cached easily because the contents
@@ -40,7 +42,7 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
                                      struct mca_coll_base_module_2_1_0_t *module) {
   int peer, rank, maxr, p, res, count;
   MPI_Aint ext;
-  ptrdiff_t gap, span;
+  ptrdiff_t gap, span, span_align;
   char *sbuf, inplace;
   NBC_Schedule *schedule;
   NBC_Handle *handle;
@@ -84,14 +86,15 @@ int ompi_coll_libnbc_ireduce_scatter(const void* sendbuf, void* recvbuf, const i
   maxr = (int) ceil ((log((double) p) / LOG2));
 
   span = opal_datatype_span(&datatype->super, count, &gap);
-  handle->tmpbuf = malloc (span * 2);
+  span_align = OPAL_ALIGN(span, datatype->super.align, ptrdiff_t);
+  handle->tmpbuf = malloc (span_align + span);
   if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
     NBC_Return_handle (handle);
     return OMPI_ERR_OUT_OF_RESOURCE;
   }
 
   rbuf = (char *)(-gap);
-  lbuf = (char *)(span - gap);
+  lbuf = (char *)(span_align - gap);
 
   schedule = OBJ_NEW(NBC_Schedule);
   if (OPAL_UNLIKELY(NULL == schedule)) {
@@ -205,7 +208,7 @@ int ompi_coll_libnbc_ireduce_scatter_inter (const void* sendbuf, void* recvbuf,
                                             struct mca_coll_base_module_2_1_0_t *module) {
   int rank, res, count, lsize, rsize;
   MPI_Aint ext;
-  ptrdiff_t gap, span;
+  ptrdiff_t gap, span, span_align;
   NBC_Schedule *schedule;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
@@ -226,14 +229,15 @@ int ompi_coll_libnbc_ireduce_scatter_inter (const void* sendbuf, void* recvbuf,
   }
 
   span = opal_datatype_span(&datatype->super, count, &gap);
+  span_align = OPAL_ALIGN(span, datatype->super.align, ptrdiff_t);
 
   res = NBC_Init_handle(comm, &handle, libnbc_module);
   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
     return res;
   }
 
   if (count > 0) {
-    handle->tmpbuf = malloc (2 * span);
+    handle->tmpbuf = malloc (span_align + span);
     if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
       NBC_Return_handle (handle);
       return OMPI_ERR_OUT_OF_RESOURCE;
@@ -259,7 +263,7 @@ int ompi_coll_libnbc_ireduce_scatter_inter (const void* sendbuf, void* recvbuf,
   if (0 == rank) {
     char *lbuf, *rbuf;
     lbuf = (char *)(-gap);
-    rbuf = (char *)(span-gap);
+    rbuf = (char *)(span_align-gap);
     res = NBC_Sched_recv (lbuf, true, count, datatype, 0, schedule, true);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       NBC_Return_handle (handle);

ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c

@@ -14,6 +14,8 @@
  * Author(s): Torsten Hoefler <[email protected]>
  *
  */
+#include "opal/include/opal/align.h"
+
 #include "nbc_internal.h"
 
 /* an reduce_csttare schedule can not be cached easily because the contents
@@ -75,18 +77,20 @@ int ompi_coll_libnbc_ireduce_scatter_block(const void* sendbuf, void* recvbuf, i
 
   if (0 < count) {
     char *rbuf, *lbuf, *buf;
+    ptrdiff_t span_align;
 
     span = opal_datatype_span(&datatype->super, count, &gap);
-    handle->tmpbuf = malloc (2*span);
+    span_align = OPAL_ALIGN(span, datatype->super.align, ptrdiff_t);
+    handle->tmpbuf = malloc (span_align + span);
     if (NULL == handle->tmpbuf) {
       OMPI_COLL_LIBNBC_REQUEST_RETURN(handle);
       OBJ_RELEASE(schedule);
       return OMPI_ERR_OUT_OF_RESOURCE;
     }
 
     rbuf = (void *)(-gap);
-    lbuf = (char *)(span - gap);
-    redbuf = (char *) handle->tmpbuf + span - gap;
+    lbuf = (char *)(span_align - gap);
+    redbuf = (char *) handle->tmpbuf + span_align - gap;
 
     /* copy data to redbuf if we only have a single node */
     if ((p == 1) && !inplace) {
@@ -206,7 +210,7 @@ int ompi_coll_libnbc_ireduce_scatter_block_inter(const void *sendbuf, void *recv
 						 ompi_request_t **request, struct mca_coll_base_module_2_1_0_t *module) {
   int rank, res, count, lsize, rsize;
   MPI_Aint ext;
-  ptrdiff_t gap, span;
+  ptrdiff_t gap, span, span_align;
   NBC_Schedule *schedule;
   NBC_Handle *handle;
   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
@@ -229,9 +233,10 @@ int ompi_coll_libnbc_ireduce_scatter_block_inter(const void *sendbuf, void *recv
   count = rcount * lsize;
 
   span = opal_datatype_span(&dtype->super, count, &gap);
+  span_align = OPAL_ALIGN(span, dtype->super.align, ptrdiff_t);
 
   if (count > 0) {
-    handle->tmpbuf = malloc (2 * span);
+    handle->tmpbuf = malloc (span_align + span);
     if (NULL == handle->tmpbuf) {
       NBC_Return_handle (handle);
       return OMPI_ERR_OUT_OF_RESOURCE;
@@ -257,7 +262,7 @@ int ompi_coll_libnbc_ireduce_scatter_block_inter(const void *sendbuf, void *recv
   if (0 == rank) {
     char *lbuf, *rbuf;
     lbuf = (char *)(-gap);
-    rbuf = (char *)(span-gap);
+    rbuf = (char *)(span_align-gap);
     res = NBC_Sched_recv (lbuf, true, count, dtype, 0, schedule, true);
     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
       NBC_Return_handle (handle);