ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

More endian-related cleanups:

Просмотр исходного кода 
- moved hton64 and ntoh64 from the bunch of places it had been copied
    into one header file
  - properly set and use the btl_tcp's nbo option to put things in
    network byte order on the wire if both sides don't have the same
    endianness
  - Put the OB1 PML's headers (with a couple exceptions I need to discuss
    with Tim) in network byte order on the wire if both sides don't have
    the same endianness
  - since it was needed for the TCP BTL, move the orte_process_name_t
    HTON and NTOH macros from the TCP OOB to ns_types.h

This commit was SVN r9145.
Этот коммит содержится в:
Brian Barrett 2006-02-26 00:45:54 +00:00
родитель 88448cf40a
Коммит 285581dff2
17 изменённых файлов: 345 добавлений и 211 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

3
ompi/mca/btl/btl.h
Просмотреть файл

@ -219,6 +219,9 @@ struct mca_btl_base_header_t{
};
typedef struct mca_btl_base_header_t mca_btl_base_header_t;
#define MCA_BTL_BASE_HEADER_HTON(hdr)
#define MCA_BTL_BASE_HEADER_NTOH(hdr)
/*
* BTL component interface functions and datatype.
*/

3
ompi/mca/btl/tcp/btl_tcp.c
Просмотреть файл

@ -390,6 +390,7 @@ int mca_btl_tcp_send(
frag->hdr.base.tag = tag;
frag->hdr.type = MCA_BTL_TCP_HDR_TYPE_SEND;
frag->hdr.count = 0;
if (endpoint->endpoint_nbo) MCA_BTL_TCP_HDR_HTON(frag->hdr);
return mca_btl_tcp_endpoint_send(endpoint,frag);
}
@ -431,6 +432,7 @@ int mca_btl_tcp_put(
frag->hdr.base.tag = MCA_BTL_TAG_BTL;
frag->hdr.type = MCA_BTL_TCP_HDR_TYPE_PUT;
frag->hdr.count = frag->base.des_dst_cnt;
if (endpoint->endpoint_nbo) MCA_BTL_TCP_HDR_HTON(frag->hdr);
return mca_btl_tcp_endpoint_send(endpoint,frag);
}
@ -466,6 +468,7 @@ int mca_btl_tcp_get(
frag->hdr.base.tag = MCA_BTL_TAG_BTL;
frag->hdr.type = MCA_BTL_TCP_HDR_TYPE_GET;
frag->hdr.count = frag->base.des_src_cnt;
if (endpoint->endpoint_nbo) MCA_BTL_TCP_HDR_HTON(frag->hdr);
return mca_btl_tcp_endpoint_send(endpoint,frag);
}

1
ompi/mca/btl/tcp/btl_tcp_component.c
Просмотреть файл

@ -601,6 +601,7 @@ static void mca_btl_tcp_component_recv_handler(int sd, short flags, void* user)
close(sd);
return;
}
OMPI_PROCESS_NAME_NTOH(guid);
/* now set socket up to be non-blocking */
if((flags = fcntl(sd, F_GETFL, 0)) < 0) {

8
ompi/mca/btl/tcp/btl_tcp_endpoint.c
Просмотреть файл

@ -277,8 +277,11 @@ static int mca_btl_tcp_endpoint_send_connect_ack(mca_btl_base_endpoint_t* btl_en
{
/* send process identifier to remote endpoint */
mca_btl_tcp_proc_t* btl_proc = mca_btl_tcp_proc_local();
if(mca_btl_tcp_endpoint_send_blocking(btl_endpoint, &btl_proc->proc_name, sizeof(btl_proc->proc_name)) !=
sizeof(btl_proc->proc_name)) {
orte_process_name_t guid = btl_proc->proc_name;
OMPI_PROCESS_NAME_HTON(guid);
if(mca_btl_tcp_endpoint_send_blocking(btl_endpoint, &guid, sizeof(guid)) !=
sizeof(guid)) {
return OMPI_ERR_UNREACH;
}
return OMPI_SUCCESS;
@ -438,6 +441,7 @@ static int mca_btl_tcp_endpoint_recv_connect_ack(mca_btl_base_endpoint_t* btl_en
if((mca_btl_tcp_endpoint_recv_blocking(btl_endpoint, &guid, sizeof(orte_process_name_t))) != sizeof(orte_process_name_t)) {
return OMPI_ERR_UNREACH;
}
OMPI_PROCESS_NAME_NTOH(guid);
/* compare this to the expected values */
if(memcmp(&btl_proc->proc_name, &guid, sizeof(orte_process_name_t)) != 0) {

1
ompi/mca/btl/tcp/btl_tcp_frag.c
Просмотреть файл

@ -206,6 +206,7 @@ bool mca_btl_tcp_frag_recv(mca_btl_tcp_frag_t* frag, int sd)
/* read header */
if(frag->iov_cnt == 0) {
if (btl_endpoint->endpoint_nbo) MCA_BTL_TCP_HDR_NTOH(frag->hdr);
switch(frag->hdr.type) {
case MCA_BTL_TCP_HDR_TYPE_SEND:
if(frag->iov_idx == 1 && frag->hdr.size) {

13
ompi/mca/btl/tcp/btl_tcp_hdr.h
Просмотреть файл

@ -23,6 +23,7 @@
#include "ompi_config.h"
#include "ompi/mca/btl/base/base.h"
#include "btl_tcp.h"
#include "opal/types.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
@ -45,6 +46,18 @@ struct mca_btl_tcp_hdr_t {
};
typedef struct mca_btl_tcp_hdr_t mca_btl_tcp_hdr_t;
#define MCA_BTL_TCP_HDR_HTON(hdr) \
do { \
hdr.count = htons(hdr.count); \
hdr.size = hton64(hdr.size); \
} while (0)
#define MCA_BTL_TCP_HDR_NTOH(hdr) \
do { \
hdr.count = ntohs(hdr.count); \
hdr.size = ntoh64(hdr.size); \
} while (0)
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif

15
ompi/mca/btl/tcp/btl_tcp_proc.c
Просмотреть файл

@ -28,6 +28,7 @@
#include "orte/class/orte_proc_table.h"
#include "ompi/mca/btl/base/btl_base_error.h"
#include "ompi/mca/pml/base/pml_base_module_exchange.h"
#include "ompi/datatype/dt_arch.h"
#include "btl_tcp.h"
#include "btl_tcp_proc.h"
@ -172,6 +173,20 @@ int mca_btl_tcp_proc_insert(
size_t i;
unsigned long net1;
#ifndef WORDS_BIGENDIAN
/* if we are little endian and our peer is not so lucky, then we
need to put all information sent to him in big endian (aka
Network Byte Order) and expect all information received to
be in NBO. Since big endian machines always send and receive
in NBO, we don't care so much about that case. */
if (btl_proc->proc_ompi->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
printf("setting BIGENDIAN\n");
btl_endpoint->endpoint_nbo = true;
} else {
printf("not setting big endian\n");
}
#endif
/* insert into endpoint array */
btl_endpoint->endpoint_proc = btl_proc;
btl_proc->proc_endpoints[btl_proc->proc_endpoint_count++] = btl_endpoint;

40
ompi/mca/osc/pt2pt/osc_pt2pt_header.h
Просмотреть файл

@ -21,6 +21,7 @@
#include <netinet/in.h>
#endif
#include "opal/types.h"
#define OMPI_OSC_PT2PT_HDR_PUT 0x0001
#define OMPI_OSC_PT2PT_HDR_ACC 0x0002
@ -134,43 +135,4 @@ typedef struct ompi_osc_pt2pt_control_header_t ompi_osc_pt2pt_control_header_t;
(hdr).hdr_value[1] = ntohl((hdr).hdr_value[1]); \
} while (0)
/*
* Convert a 64 bit value to network byte order.
*/
static inline uint64_t hton64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = htonl(w.l[1]);
r.l[1] = htonl(w.l[0]);
return r.ll;
}
/*
* Convert a 64 bit value from network to host byte order.
*/
static inline uint64_t ntoh64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = ntohl(w.l[1]);
r.l[1] = ntohl(w.l[0]);
return r.ll;
}
#endif /* OMPI_MCA_OSC_PT2PT_HDR_H */

42
ompi/mca/pml/dr/pml_dr_hdr.h
Просмотреть файл

@ -29,6 +29,8 @@
#include <netinet/in.h>
#endif
#include "opal/types.h"
#define MCA_PML_DR_HDR_TYPE_MATCH 1
#define MCA_PML_DR_HDR_TYPE_RNDV 2
#define MCA_PML_DR_HDR_TYPE_ACK 3
@ -39,46 +41,6 @@
#define MCA_PML_DR_HDR_FLAGS_MATCH 4 /* is the ack in response to a match */
/*
* Convert a 64 bit value to network byte order.
*/
static inline uint64_t hton64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = htonl(w.l[1]);
r.l[1] = htonl(w.l[0]);
return r.ll;
}
/*
* Convert a 64 bit value from network to host byte order.
*/
static inline uint64_t ntoh64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = ntohl(w.l[1]);
r.l[1] = ntohl(w.l[0]);
return r.ll;
}
/**
* Common hdr attributes - must be first element in each hdr type
*/

57
ompi/mca/pml/ob1/pml_ob1_hdr.h
Просмотреть файл

@ -29,6 +29,8 @@
#include <netinet/in.h>
#endif
#include "opal/types.h"
#define MCA_PML_OB1_HDR_TYPE_MATCH 1
#define MCA_PML_OB1_HDR_TYPE_RNDV 2
#define MCA_PML_OB1_HDR_TYPE_RGET 3
@ -45,47 +47,6 @@
#define MCA_PML_OB1_HDR_FLAGS_PIN 4 /* is user buffer pinned */
#define MCA_PML_OB1_HDR_FLAGS_CONTIG 8 /* is user buffer contiguous */
/*
* Convert a 64 bit value to network byte order.
*/
static inline uint64_t hton64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = htonl(w.l[1]);
r.l[1] = htonl(w.l[0]);
return r.ll;
}
/*
* Convert a 64 bit value from network to host byte order.
*/
static inline uint64_t ntoh64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = ntohl(w.l[1]);
r.l[1] = ntohl(w.l[0]);
return r.ll;
}
/**
* Common hdr attributes - must be first element in each hdr type
*/
@ -203,13 +164,13 @@ typedef struct mca_pml_ob1_ack_hdr_t mca_pml_ob1_ack_hdr_t;
#define MCA_PML_OB1_ACK_HDR_NTOH(h) \
do { \
MCA_PML_OB1_COMMON_HDR_NTOH(h.hdr_common); \
(h).hdr_dst_size = ntoh64((h).hdr_dst_size); \
(h).hdr_rdma_offset = ntoh64((h).hdr_rdma_offset); \
} while (0)
#define MCA_PML_OB1_ACK_HDR_HTON(h) \
do { \
MCA_PML_OB1_COMMON_HDR_HTON((h).hdr_common); \
(h).hdr_dst_size = hton64((h).hdr_dst_size); \
(h).hdr_rdma_offset = hton64((h).hdr_rdma_offset); \
} while (0)
/**
@ -236,6 +197,16 @@ struct mca_pml_ob1_fin_hdr_t {
};
typedef struct mca_pml_ob1_fin_hdr_t mca_pml_ob1_fin_hdr_t;
#define MCA_PML_OB1_FIN_HDR_NTOH(h) \
do { \
MCA_PML_OB1_COMMON_HDR_NTOH(h.hdr_common); \
} while (0)
#define MCA_PML_OB1_FIN_HDR_HTON(h) \
do { \
MCA_PML_OB1_COMMON_HDR_HTON((h).hdr_common); \
} while (0)
/**
* Union of defined hdr types.
*/

53
ompi/mca/pml/ob1/pml_ob1_recvfrag.c
Просмотреть файл

@ -33,6 +33,7 @@
#include "pml_ob1_recvreq.h"
#include "pml_ob1_sendreq.h"
#include "pml_ob1_hdr.h"
#include "ompi/datatype/dt_arch.h"
@ -68,17 +69,45 @@ void mca_pml_ob1_recv_frag_callback(
return;
}
/* hdr_type and hdr_flags are uint8_t, so no endian problems */
switch(hdr->hdr_common.hdr_type) {
case MCA_PML_OB1_HDR_TYPE_MATCH:
{
#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
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
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
/* BWB - FIX ME - Tim, what do I do with 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 = (mca_pml_ob1_send_request_t*)
mca_pml_ob1_send_request_t* sendreq;
#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
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;
sendreq->req_rdma_offset = hdr->hdr_ack.hdr_rdma_offset;
@ -87,21 +116,37 @@ void mca_pml_ob1_recv_frag_callback(
}
case MCA_PML_OB1_HDR_TYPE_FRAG:
{
mca_pml_ob1_recv_request_t* recvreq = (mca_pml_ob1_recv_request_t*)
mca_pml_ob1_recv_request_t* recvreq;
#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
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;
}
case MCA_PML_OB1_HDR_TYPE_PUT:
{
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)
mca_pml_ob1_send_request_t* sendreq;
#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* BWB - FIX ME - Tim, what do I do with rdma headers? */
#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;
}
case MCA_PML_OB1_HDR_TYPE_FIN:
{
mca_btl_base_descriptor_t* rdma = (mca_btl_base_descriptor_t*)
mca_btl_base_descriptor_t* rdma;
#if !defined(WORDS_BIGENDIAN) && OMPI_ENABLE_HETEROGENEOUS_SUPPORT
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, OMPI_SUCCESS);
break;

38
ompi/mca/pml/ob1/pml_ob1_recvreq.c
Просмотреть файл

@ -29,7 +29,8 @@
#include "pml_ob1_rdmafrag.h"
#include "ompi/mca/bml/base/base.h"
#include "orte/mca/errmgr/errmgr.h"
#include "ompi/datatype/dt_arch.h"
static mca_pml_ob1_recv_frag_t* mca_pml_ob1_recv_request_match_specific_proc(
mca_pml_ob1_recv_request_t* request, mca_pml_ob1_comm_proc_t* proc);
@ -264,6 +265,18 @@ static void mca_pml_ob1_recv_request_ack(
ack->hdr_dst_req.pval = recvreq;
ack->hdr_rdma_offset = recvreq->req_rdma_offset;
#ifdef WORDS_BIGENDIAN
ack->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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) {
ack->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
MCA_PML_OB1_ACK_HDR_HTON(*ack);
}
#endif
/* initialize descriptor */
des->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY;
des->des_cbfunc = mca_pml_ob1_ctl_completion;
@ -350,6 +363,18 @@ static void mca_pml_ob1_rget_completion(
hdr->hdr_common.hdr_flags = 0;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_FIN;
hdr->hdr_des = frag->rdma_hdr.hdr_rget.hdr_des;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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_FIN_HDR_HTON(*hdr);
}
#endif
/* queue request */
rc = mca_bml_base_send(
@ -716,6 +741,17 @@ void mca_pml_ob1_recv_request_schedule(mca_pml_ob1_recv_request_t* recvreq)
hdr->hdr_seg_cnt = dst->des_dst_cnt;
memcpy(hdr->hdr_segs, dst->des_dst, dst->des_dst_cnt * sizeof(mca_btl_base_segment_t));
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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) {
/* BWB - FIX ME - TIM, what do we do here? */
}
#endif
/* update request state */
recvreq->req_rdma_offset += size;
OPAL_THREAD_ADD_SIZE_T(&recvreq->req_pipeline_depth,1);

98
ompi/mca/pml/ob1/pml_ob1_sendreq.c
Просмотреть файл

@ -33,6 +33,7 @@
#include "pml_ob1_recvreq.h"
#include "pml_ob1_endpoint.h"
#include "ompi/mca/bml/base/base.h"
#include "ompi/datatype/dt_arch.h"
static int mca_pml_ob1_send_request_fini(struct ompi_request_t** request)
@ -336,6 +337,18 @@ 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;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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
/* update lengths */
segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t) + max_data;
sendreq->req_send_offset = max_data;
@ -437,6 +450,18 @@ int mca_pml_ob1_send_request_start_copy(
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_seq = sendreq->req_send.req_base.req_sequence;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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
/* update lengths */
segment->seg_len = sizeof(mca_pml_ob1_match_hdr_t) + max_data;
sendreq->req_send_offset = max_data;
@ -497,6 +522,18 @@ int mca_pml_ob1_send_request_start_prepare(
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_seq = sendreq->req_send.req_base.req_sequence;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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
/* short message */
descriptor->des_cbfunc = mca_pml_ob1_match_completion_free;
@ -578,6 +615,19 @@ int mca_pml_ob1_send_request_start_rdma(
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
hdr->hdr_rget.hdr_des.pval = src;
hdr->hdr_rget.hdr_seg_cnt = src->des_src_cnt;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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) {
/* BWB - FIX ME - Ask Tim what we are supposed to do in this case */
return OMPI_ERR_NOT_SUPPORTED;
}
#endif
for(i=0; i<src->des_src_cnt; i++)
hdr->hdr_rget.hdr_segs[i] = src->des_src[i];
des->des_cbfunc = mca_pml_ob1_ctl_completion;
@ -605,6 +655,18 @@ 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;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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
/* update lengths with number of bytes actually packed */
segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t);
sendreq->req_send_offset = 0;
@ -674,6 +736,18 @@ int mca_pml_ob1_send_request_start_rndv(
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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
/* first fragment of a long message */
des->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY;
des->des_cbdata = sendreq;
@ -796,6 +870,18 @@ int mca_pml_ob1_send_request_schedule(mca_pml_ob1_send_request_t* sendreq)
hdr->hdr_src_req.pval = sendreq;
hdr->hdr_dst_req = sendreq->req_recv;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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
/* update state */
sendreq->req_send_offset += size;
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_pipeline_depth,1);
@ -897,6 +983,18 @@ static void mca_pml_ob1_put_completion(
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_FIN;
hdr->hdr_des = frag->rdma_hdr.hdr_rdma.hdr_des;
#ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
/* 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_FIN_HDR_HTON(*hdr);
}
#endif
/* queue request */
rc = mca_bml_base_send(
bml_btl,

18
ompi/mca/pml/ob1/pml_ob1_sendreq.h
Просмотреть файл

@ -29,6 +29,7 @@
#include "pml_ob1_hdr.h"
#include "pml_ob1_rdma.h"
#include "ompi/datatype/convertor.h"
#include "ompi/datatype/dt_arch.h"
#include "ompi/mca/bml/bml.h"
#if defined(c_plusplus) || defined(__cplusplus)
@ -108,6 +109,22 @@ OBJ_CLASS_DECLARATION(mca_pml_ob1_send_request_t);
* Start a send request.
*/
#ifdef WORDS_BIGENDIAN
#define MCA_PML_OB1_SEND_REQUEST_START_MATCH_FIXUP(sendreq, hdr) \
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#elif OMPI_ENABLE_HETEROGENEOUS_SUPPORT
#define MCA_PML_OB1_SEND_REQUEST_START_MATCH_FIXUP(sendreq, hdr) \
do { \
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); \
} \
} while (0)
#else
#define MCA_PML_OB1_SEND_REQUEST_START_MATCH_FIXUP(sendreq, hdr)
#endif
#define MCA_PML_OB1_SEND_REQUEST_START(sendreq, rc) \
do { \
mca_pml_ob1_comm_t* comm = sendreq->req_send.req_base.req_comm->c_pml_comm; \
@ -173,6 +190,7 @@ do {
hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; \
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; \
hdr->hdr_match.hdr_seq = sendreq->req_send.req_base.req_sequence; \
MCA_PML_OB1_SEND_REQUEST_START_MATCH_FIXUP(sendreq, hdr); \
\
/* short message */ \
descriptor->des_cbfunc = mca_pml_ob1_match_completion_cache; \

38
opal/include/opal/types.h
Просмотреть файл

@ -103,4 +103,42 @@ typedef socklen_t ompi_socklen_t;
typedef int ompi_socklen_t;
#endif
/*
* Convert a 64 bit value to network byte order.
*/
static inline uint64_t hton64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = htonl(w.l[1]);
r.l[1] = htonl(w.l[0]);
return r.ll;
}
/*
* Convert a 64 bit value from network to host byte order.
*/
static inline uint64_t ntoh64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = ntohl(w.l[1]);
r.l[1] = ntohl(w.l[0]);
return r.ll;
}
#endif

46
orte/mca/ns/ns_types.h
Просмотреть файл

@ -89,6 +89,52 @@ typedef struct orte_process_name_t orte_process_name_t;
extern orte_process_name_t orte_name_all;
#define ORTE_NAME_ALL &orte_name_all
#if SIZEOF_SIZE_T == 8
/**
* Convert process name from host to network byte order.
*
* @param name
*/
#define OMPI_PROCESS_NAME_HTON(n) \
n.cellid = hton64(n.cellid); \
n.jobid = hton64(n.jobid); \
n.vpid = hton64(n.vpid);
/**
* Convert process name from network to host byte order.
*
* @param name
*/
#define OMPI_PROCESS_NAME_NTOH(n) \
n.cellid = ntoh64(n.cellid); \
n.jobid = ntoh64(n.jobid); \
n.vpid = ntoh64(n.vpid);
#else
/**
* Convert process name from host to network byte order.
*
* @param name
*/
#define OMPI_PROCESS_NAME_HTON(n) \
n.cellid = htonl(n.cellid); \
n.jobid = htonl(n.jobid); \
n.vpid = htonl(n.vpid);
/**
* Convert process name from network to host byte order.
*
* @param name
*/
#define OMPI_PROCESS_NAME_NTOH(n) \
n.cellid = ntohl(n.cellid); \
n.jobid = ntohl(n.jobid); \
n.vpid = ntohl(n.vpid);
#endif
/** List of names for general use
*/

82
orte/mca/oob/tcp/oob_tcp.h
Просмотреть файл

@ -61,88 +61,6 @@ int mca_oob_tcp_init(void);
*/
int mca_oob_tcp_fini(void);
#if SIZEOF_SIZE_T == 8
/*
* Convert a 64 bit value to network byte order.
*/
static inline uint64_t hton64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = htonl(w.l[1]);
r.l[1] = htonl(w.l[0]);
return r.ll;
}
/*
* Convert a 64 bit value from network to host byte order.
*/
static inline uint64_t ntoh64(uint64_t val)
{
union { uint64_t ll;
uint32_t l[2];
} w, r;
/* platform already in network byte order? */
if(htonl(1) == 1L)
return val;
w.ll = val;
r.l[0] = ntohl(w.l[1]);
r.l[1] = ntohl(w.l[0]);
return r.ll;
}
/**
* Convert process name from host to network byte order.
*
* @param name
*/
#define OMPI_PROCESS_NAME_HTON(n) \
n.cellid = hton64(n.cellid); \
n.jobid = hton64(n.jobid); \
n.vpid = hton64(n.vpid);
/**
* Convert process name from network to host byte order.
*
* @param name
*/
#define OMPI_PROCESS_NAME_NTOH(n) \
n.cellid = ntoh64(n.cellid); \
n.jobid = ntoh64(n.jobid); \
n.vpid = ntoh64(n.vpid);
#else
/**
* Convert process name from host to network byte order.
*
* @param name
*/
#define OMPI_PROCESS_NAME_HTON(n) \
n.cellid = htonl(n.cellid); \
n.jobid = htonl(n.jobid); \
n.vpid = htonl(n.vpid);
/**
* Convert process name from network to host byte order.
*
* @param name
*/
#define OMPI_PROCESS_NAME_NTOH(n) \
n.cellid = ntohl(n.cellid); \
n.jobid = ntohl(n.jobid); \
n.vpid = ntohl(n.vpid);
#endif
/**
* Compare two process names for equality.
*