Better encapsulate heterogeneous arch handling in ob1.

This commit was SVN r16970.

ompi/mca/pml/ob1/pml_ob1.c

@@ -348,19 +348,7 @@ int mca_pml_ob1_send_fin( ompi_proc_t* proc,
     hdr->hdr_des.pval = hdr_des;
     hdr->hdr_fail = status;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+    ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_FIN, proc);
-#ifdef WORDS_BIGENDIAN
-    hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-    /* if we are little endian and the remote side is big endian,
-       we're responsible for making sure the data is in network byte
-       order */
-    if (proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-        MCA_PML_OB1_FIN_HDR_HTON(*hdr);
-    }
-#endif
-#endif
 
     /* queue request */
     rc = mca_bml_base_send(

ompi/mca/pml/ob1/pml_ob1_hdr.h

@@ -30,6 +30,7 @@
 #endif
 
 #include "opal/types.h"
+#include "ompi/datatype/dt_arch.h"
 
 #define MCA_PML_OB1_HDR_TYPE_MATCH     1
 #define MCA_PML_OB1_HDR_TYPE_RNDV      2
@@ -275,5 +276,88 @@ union mca_pml_ob1_hdr_t {
 };
 typedef union mca_pml_ob1_hdr_t mca_pml_ob1_hdr_t;
 
+#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+static inline __opal_attribute_always_inline__ void
+ob1_hdr_ntoh(mca_pml_ob1_hdr_t *hdr, const uint8_t hdr_type)
+{
+    if(!(hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO))
+        return;
 
+    switch(hdr_type) {
+        case MCA_PML_OB1_HDR_TYPE_MATCH:
+            MCA_PML_OB1_MATCH_HDR_NTOH(hdr->hdr_match);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_RNDV:
+            MCA_PML_OB1_RNDV_HDR_NTOH(hdr->hdr_rndv);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_RGET:
+            MCA_PML_OB1_RGET_HDR_NTOH(hdr->hdr_rget);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_ACK:
+            MCA_PML_OB1_ACK_HDR_NTOH(hdr->hdr_ack);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_FRAG:
+            MCA_PML_OB1_FRAG_HDR_NTOH(hdr->hdr_frag);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_PUT:
+            MCA_PML_OB1_RDMA_HDR_NTOH(hdr->hdr_rdma);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_FIN:
+            MCA_PML_OB1_FIN_HDR_NTOH(hdr->hdr_fin);
+            break;
+        default:
+            assert(0);
+            break;
+    }
+}
+#else
+#define ob1_hdr_ntoh(h, t) do{}while(0)
+#endif
+
+#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+#define ob1_hdr_hton(h, t, p) \
+    ob1_hdr_hton_intr((mca_pml_ob1_hdr_t*)h, t, p)
+static inline __opal_attribute_always_inline__ void
+ob1_hdr_hton_intr(mca_pml_ob1_hdr_t *hdr, const uint8_t hdr_type,
+        const ompi_proc_t *proc)
+{
+#ifdef WORDS_BIGENDIAN
+    hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
+#else
+
+    if(!(proc->proc_arch & OMPI_ARCH_ISBIGENDIAN))
+        return;
+
+    hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
+    switch(hdr_type) {
+        case MCA_PML_OB1_HDR_TYPE_MATCH:
+            MCA_PML_OB1_MATCH_HDR_HTON(hdr->hdr_match);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_RNDV:
+            MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_RGET:
+            MCA_PML_OB1_RGET_HDR_HTON(hdr->hdr_rget);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_ACK:
+            MCA_PML_OB1_ACK_HDR_HTON(hdr->hdr_ack);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_FRAG:
+            MCA_PML_OB1_FRAG_HDR_HTON(hdr->hdr_frag);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_PUT:
+            MCA_PML_OB1_RDMA_HDR_HTON(hdr->hdr_rdma);
+            break;
+        case MCA_PML_OB1_HDR_TYPE_FIN:
+            MCA_PML_OB1_FIN_HDR_HTON(hdr->hdr_fin);
+            break;
+        default:
+            assert(0);
+            break;
+    }
+#endif
+}
+#else
+#define ob1_hdr_hton(h, t, p) do{}while(0)
+#endif
 #endif

ompi/mca/pml/ob1/pml_ob1_recvfrag.c

@@ -34,7 +34,6 @@
 #include "pml_ob1_recvreq.h"
 #include "pml_ob1_sendreq.h"
 #include "pml_ob1_hdr.h"
-#include "ompi/datatype/dt_arch.h"
 #include "ompi/peruse/peruse-internal.h"
 
 OBJ_CLASS_INSTANCE( mca_pml_ob1_buffer_t,
@@ -86,42 +85,26 @@ void mca_pml_ob1_recv_frag_callback( mca_btl_base_module_t* btl,
     switch(hdr->hdr_common.hdr_type) {
     case MCA_PML_OB1_HDR_TYPE_MATCH:
         {
-#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+            ob1_hdr_ntoh(hdr, MCA_PML_OB1_HDR_TYPE_MATCH);
-            if (hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO) {
-                MCA_PML_OB1_MATCH_HDR_NTOH(hdr->hdr_match);
-            }
-#endif
             mca_pml_ob1_recv_frag_match(btl, &hdr->hdr_match, segments,des->des_dst_cnt);
             break;
         }
     case MCA_PML_OB1_HDR_TYPE_RNDV:
         {
-#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+            ob1_hdr_ntoh(hdr, MCA_PML_OB1_HDR_TYPE_RNDV);
-            if (hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO) {
-                MCA_PML_OB1_RNDV_HDR_NTOH(hdr->hdr_rndv);
-            }
-#endif
             mca_pml_ob1_recv_frag_match(btl, &hdr->hdr_match, segments,des->des_dst_cnt);
             break;
         }
     case MCA_PML_OB1_HDR_TYPE_RGET:
         {
-#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+            ob1_hdr_ntoh(hdr, MCA_PML_OB1_HDR_TYPE_RGET);
-            if (hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO) {
-                MCA_PML_OB1_RGET_HDR_NTOH(hdr->hdr_rget);
-            }
-#endif
             mca_pml_ob1_recv_frag_match(btl, &hdr->hdr_match, segments,des->des_dst_cnt);
             break;
         }
     case MCA_PML_OB1_HDR_TYPE_ACK:
         {
             mca_pml_ob1_send_request_t* sendreq;
-#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+            ob1_hdr_ntoh(hdr, MCA_PML_OB1_HDR_TYPE_ACK);
-            if (hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO) {
-                MCA_PML_OB1_ACK_HDR_NTOH(hdr->hdr_ack);
-            }
-#endif
             sendreq = (mca_pml_ob1_send_request_t*)hdr->hdr_ack.hdr_src_req.pval;
             sendreq->req_recv = hdr->hdr_ack.hdr_dst_req;
 
@@ -145,11 +128,7 @@ void mca_pml_ob1_recv_frag_callback( mca_btl_base_module_t* btl,
     case MCA_PML_OB1_HDR_TYPE_FRAG:
         {
             mca_pml_ob1_recv_request_t* recvreq;
-#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+            ob1_hdr_ntoh(hdr, MCA_PML_OB1_HDR_TYPE_FRAG);
-            if (hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO) {
-                MCA_PML_OB1_FRAG_HDR_NTOH(hdr->hdr_frag);
-            }
-#endif
             recvreq = (mca_pml_ob1_recv_request_t*)hdr->hdr_frag.hdr_dst_req.pval;
             mca_pml_ob1_recv_request_progress(recvreq,btl,segments,des->des_dst_cnt);
             break;
@@ -157,11 +136,7 @@ void mca_pml_ob1_recv_frag_callback( mca_btl_base_module_t* btl,
     case MCA_PML_OB1_HDR_TYPE_PUT:
         {
             mca_pml_ob1_send_request_t* sendreq;
-#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+            ob1_hdr_ntoh(hdr, MCA_PML_OB1_HDR_TYPE_PUT);
-            if (hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO) {
-                MCA_PML_OB1_RDMA_HDR_NTOH(hdr->hdr_rdma);
-            }
-#endif
             sendreq = (mca_pml_ob1_send_request_t*)hdr->hdr_rdma.hdr_req.pval;
             mca_pml_ob1_send_request_put(sendreq,btl,&hdr->hdr_rdma);
             break;
@@ -169,11 +144,7 @@ void mca_pml_ob1_recv_frag_callback( mca_btl_base_module_t* btl,
     case MCA_PML_OB1_HDR_TYPE_FIN:
         {
             mca_btl_base_descriptor_t* rdma;
-#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+            ob1_hdr_ntoh(hdr, MCA_PML_OB1_HDR_TYPE_FIN);
-            if (hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO) {
-                MCA_PML_OB1_FIN_HDR_NTOH(hdr->hdr_fin);
-            }
-#endif
             rdma = (mca_btl_base_descriptor_t*)hdr->hdr_fin.hdr_des.pval;
             rdma->des_cbfunc(btl, NULL, rdma,
                     hdr->hdr_fin.hdr_fail ? OMPI_ERROR : OMPI_SUCCESS);

ompi/mca/pml/ob1/pml_ob1_recvreq.c

@@ -203,19 +203,7 @@ int mca_pml_ob1_recv_request_ack_send_btl(
     ack->hdr_dst_req.pval = hdr_dst_req;
     ack->hdr_send_offset = hdr_send_offset;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+    ob1_hdr_hton(ack, MCA_PML_OB1_HDR_TYPE_ACK, proc);
-#ifdef WORDS_BIGENDIAN
-    ack->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-    /* if we are little endian and the remote side is big endian,
-       we're responsible for making sure the data is in network byte
-       order */
-    if (proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
-        ack->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-        MCA_PML_OB1_ACK_HDR_HTON(*ack);
-    }
-#endif
-#endif
 
     /* initialize descriptor */
     des->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY;
@@ -704,20 +692,8 @@ int mca_pml_ob1_recv_request_schedule_once(
 
         if(!recvreq->req_ack_sent)
             recvreq->req_ack_sent = true;
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+        ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_PUT,
-#ifdef WORDS_BIGENDIAN
+                recvreq->req_recv.req_base.req_proc);
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-        /* if we are little endian and the remote side is big endian,
-           we're responsible for making sure the data is in network byte
-           order */
-        if(recvreq->req_recv.req_base.req_proc->proc_arch &
-                OMPI_ARCH_ISBIGENDIAN) {
-            hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-            MCA_PML_OB1_RDMA_HDR_HTON(*hdr);
-        }
-#endif
-#endif
 
         PERUSE_TRACE_COMM_OMPI_EVENT( PERUSE_COMM_REQ_XFER_CONTINUE,
                                       &(recvreq->req_recv.req_base), size,

ompi/mca/pml/ob1/pml_ob1_sendreq.c

@@ -31,7 +31,6 @@
 #include "pml_ob1_rdmafrag.h"
 #include "pml_ob1_recvreq.h"
 #include "ompi/mca/bml/base/base.h"
-#include "ompi/datatype/dt_arch.h"
 
 OBJ_CLASS_INSTANCE(mca_pml_ob1_send_range_t, ompi_free_list_item_t,
         NULL, NULL);
@@ -361,19 +360,8 @@ int mca_pml_ob1_send_request_start_buffered(
     hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
     hdr->hdr_rndv.hdr_src_req.pval = sendreq;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+    ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_RNDV,
-#ifdef WORDS_BIGENDIAN
+            sendreq->req_send.req_base.req_proc);
-    hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-    /* if we are little endian and the remote side is big endian,
-       we're responsible for making sure the data is in network byte
-       order */
-    if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-        MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv);
-    }
-#endif
-#endif
 
     /* update lengths */
     segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t) + max_data;
@@ -470,19 +458,8 @@ int mca_pml_ob1_send_request_start_copy( mca_pml_ob1_send_request_t* sendreq,
     hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
     hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+    ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_MATCH,
-#ifdef WORDS_BIGENDIAN
+            sendreq->req_send.req_base.req_proc);
-    hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-    /* if we are little endian and the remote side is big endian,
-       we're responsible for making sure the data is in network byte
-       order */
-    if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-        MCA_PML_OB1_MATCH_HDR_HTON(hdr->hdr_match);
-    }
-#endif
-#endif
 
     /* update lengths */
     segment->seg_len = sizeof(mca_pml_ob1_match_hdr_t) + max_data;
@@ -550,19 +527,8 @@ int mca_pml_ob1_send_request_start_prepare( mca_pml_ob1_send_request_t* sendreq,
     hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
     hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+    ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_MATCH,
-#ifdef WORDS_BIGENDIAN
+            sendreq->req_send.req_base.req_proc);
-    hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-    /* if we are little endian and the remote side is big endian,
-       we're responsible for making sure the data is in network byte
-       order */
-    if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-        MCA_PML_OB1_MATCH_HDR_HTON(hdr->hdr_match);
-    }
-#endif
-#endif
 
     /* short message */
     descriptor->des_cbfunc = mca_pml_ob1_match_completion_free;
@@ -654,19 +620,8 @@ int mca_pml_ob1_send_request_start_rdma(
         hdr->hdr_rget.hdr_des.pval = src;
         hdr->hdr_rget.hdr_seg_cnt = src->des_src_cnt;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+        ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_RGET,
-#ifdef WORDS_BIGENDIAN
+                sendreq->req_send.req_base.req_proc);
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-        /* if we are little endian and the remote side is big endian,
-           we're responsible for making sure the data is in network byte
-           order */
-        if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
-            hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-            MCA_PML_OB1_RGET_HDR_HTON(hdr->hdr_rget);
-        }
-#endif
-#endif
 
         for( i = 0; i < src->des_src_cnt; i++ ) {
             hdr->hdr_rget.hdr_segs[i].seg_addr.lval = ompi_ptr_ptol(src->des_src[i].seg_addr.pval);
@@ -712,19 +667,8 @@ int mca_pml_ob1_send_request_start_rdma(
         hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
         hdr->hdr_rndv.hdr_src_req.pval = sendreq;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+        ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_RNDV,
-#ifdef WORDS_BIGENDIAN
+                sendreq->req_send.req_base.req_proc);
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-        /* if we are little endian and the remote side is big endian,
-           we're responsible for making sure the data is in network byte
-           order */
-        if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
-            hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-            MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv);
-        }
-#endif
-#endif
 
         /* update lengths with number of bytes actually packed */
         segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t);
@@ -797,19 +741,8 @@ int mca_pml_ob1_send_request_start_rndv( mca_pml_ob1_send_request_t* sendreq,
     hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
     hdr->hdr_rndv.hdr_src_req.pval = sendreq;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+    ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_RNDV,
-#ifdef WORDS_BIGENDIAN
+            sendreq->req_send.req_base.req_proc);
-    hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-    /* if we are little endian and the remote side is big endian,
-       we're responsible for making sure the data is in network byte
-       order */
-    if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-        MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv);
-    }
-#endif
-#endif
 
     /* first fragment of a long message */
     des->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY;
@@ -1010,22 +943,8 @@ cannot_pack:
         hdr->hdr_src_req.pval = sendreq;
         hdr->hdr_dst_req = sendreq->req_recv;
 
-#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
+        ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_FRAG,
-#ifdef WORDS_BIGENDIAN
+                sendreq->req_send.req_base.req_proc);
-        hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-#else
-        /*
-         * if we are little endian and the remote side is big endian,
-         *  we're responsible for making sure the data is in network byte
-         *  order
-         */
-        if(sendreq->req_send.req_base.req_proc->proc_arch &
-                OMPI_ARCH_ISBIGENDIAN) {
-            hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
-            MCA_PML_OB1_FRAG_HDR_HTON(*hdr);
-        }
-#endif
-#endif
 
 #if OMPI_WANT_PERUSE
          PERUSE_TRACE_COMM_OMPI_EVENT(PERUSE_COMM_REQ_XFER_CONTINUE,

ompi/mca/pml/ob1/pml_ob1_sendreq.h

@@ -27,7 +27,6 @@
 #include "pml_ob1_rdma.h"
 #include "pml_ob1_rdmafrag.h"
 #include "ompi/datatype/convertor.h"
-#include "ompi/datatype/dt_arch.h"
 #include "ompi/mca/bml/bml.h" 
 
 BEGIN_C_DECLS