SCALE-RM
Macros | Typedefs | Functions | Variables
scale_rdma.c File Reference
#include <stdlib.h>
#include "rdma.h"
#include <mpi.h>
#include <mpi-ext.h>
#include <stdio.h>
Include dependency graph for scale_rdma.c:

Go to the source code of this file.

Macros

#define WEST   0
 
#define NORTH   1
 
#define EAST   2
 
#define SOUTH   3
 
#define BEARING_CNT   4
 
#define offset(Z, X, Y)   (sizeof(var_t)*((Z)+(X)*(KA)+(Y)*(KA)*(IA)))
 
#define RDMA_TAG_NUM_MAX   15
 
#define RDMA_TAG   0
 
#define RDMA_TAG_TAIL   1
 
#define LOCAL_RECV_READY   0
 
#define REMOTE_RECV_READY   1
 
#define LOCAL_PUT_DONE   2
 
#define REMOTE_PUT_DONE   3
 
#define LOCAL_RECV_DONE   3
 
#define FLAG_CNT   4
 
#define sidx(DIR, FLAG)   ((DIR)+(BEARING_CNT)*(FLAG))
 
#define soffset(DIR, FLAG)   (sizeof(sf_t)*sidx(DIR,FLAG))
 
#define FALSE   0
 
#define TRUE   1
 
#define RDMA_PUT_ID_MAX   0xFFFFFFFF
 
#define FJMPI_RDMA_PUT_FLAGS_TO_NORTH   (FJMPI_RDMA_LOCAL_NIC0 | FJMPI_RDMA_REMOTE_NIC2 | FJMPI_RDMA_PATH0 )
 
#define FJMPI_RDMA_PUT_FLAGS_TO_SOUTH   (FJMPI_RDMA_LOCAL_NIC2 | FJMPI_RDMA_REMOTE_NIC0 | FJMPI_RDMA_PATH0 )
 
#define FJMPI_RDMA_PUT_FLAGS_TO_EAST   (FJMPI_RDMA_LOCAL_NIC3 | FJMPI_RDMA_REMOTE_NIC1 | FJMPI_RDMA_PATH0 )
 
#define FJMPI_RDMA_PUT_FLAGS_TO_WEST   (FJMPI_RDMA_LOCAL_NIC1 | FJMPI_RDMA_REMOTE_NIC3 | FJMPI_RDMA_PATH0 )
 

Typedefs

typedef uint64_t sf_t
 

Functions

void rdma_setup_ (const int32_t *COMM_vsize_max_in, const int32_t *IA_in, const int32_t *JA_in, const int32_t *KA_in, const int32_t *IHALO_in, const int32_t *JHALO_in, const int32_t *IS_in, const int32_t *IE_in, const int32_t *JS_in, const int32_t *JE_in, const int32_t *RANK_W_in, const int32_t *RANK_N_in, const int32_t *RANK_E_in, const int32_t *RANK_S_in)
 
void set_rdma_variable_ (const var_t *var, const int32_t *vid)
 
void rdma_put_ (const int32_t *vid, const int32_t *num)
 
void rdma_put8_ (const int32_t *vid, const int32_t *num)
 

Variables

int32_t COMM_vsize_max
 
int32_t IA
 
int32_t JA
 
int32_t KA
 
int32_t IHALO
 
int32_t JHALO
 
int32_t IS
 
int32_t IE
 
int32_t JS
 
int32_t JE
 
size_t datasize_NS
 
size_t datasize_WE
 
int32_t RANK_W
 
int32_t RANK_N
 
int32_t RANK_E
 
int32_t RANK_S
 
int32_t memid_cnt
 
int32_t * memid
 
uint64_t * lvar
 
uint64_t ** rvar
 
volatile sf_t * status_flag
 
int32_t memid_sf
 
uint64_t local_sf
 
uint64_t * remote_sf
 
int64_t * sending
 
volatile sf_t rdma_put_id
 

Macro Definition Documentation

◆ WEST

#define WEST   0

Private parameters & variables

Definition at line 14 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ NORTH

#define NORTH   1

Definition at line 15 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ EAST

#define EAST   2

Definition at line 16 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ SOUTH

#define SOUTH   3

Definition at line 17 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ BEARING_CNT

#define BEARING_CNT   4

Definition at line 18 of file scale_rdma.c.

Referenced by rdma_setup_().

◆ offset

#define offset (   Z,
  X,
  Y 
)    (sizeof(var_t)*((Z)+(X)*(KA)+(Y)*(KA)*(IA)))

Definition at line 25 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ RDMA_TAG_NUM_MAX

#define RDMA_TAG_NUM_MAX   15

Definition at line 31 of file scale_rdma.c.

◆ RDMA_TAG

#define RDMA_TAG   0

Definition at line 32 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ RDMA_TAG_TAIL

#define RDMA_TAG_TAIL   1

Definition at line 33 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ LOCAL_RECV_READY

#define LOCAL_RECV_READY   0

Definition at line 47 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ REMOTE_RECV_READY

#define REMOTE_RECV_READY   1

Definition at line 48 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ LOCAL_PUT_DONE

#define LOCAL_PUT_DONE   2

Definition at line 49 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ REMOTE_PUT_DONE

#define REMOTE_PUT_DONE   3

Definition at line 50 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ LOCAL_RECV_DONE

#define LOCAL_RECV_DONE   3

Definition at line 51 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ FLAG_CNT

#define FLAG_CNT   4

Definition at line 52 of file scale_rdma.c.

Referenced by rdma_setup_().

◆ sidx

#define sidx (   DIR,
  FLAG 
)    ((DIR)+(BEARING_CNT)*(FLAG))

Definition at line 53 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ soffset

#define soffset (   DIR,
  FLAG 
)    (sizeof(sf_t)*sidx(DIR,FLAG))

Definition at line 54 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ FALSE

#define FALSE   0

Definition at line 57 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ TRUE

#define TRUE   1

Definition at line 58 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ RDMA_PUT_ID_MAX

#define RDMA_PUT_ID_MAX   0xFFFFFFFF

Definition at line 62 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ FJMPI_RDMA_PUT_FLAGS_TO_NORTH

#define FJMPI_RDMA_PUT_FLAGS_TO_NORTH   (FJMPI_RDMA_LOCAL_NIC0 | FJMPI_RDMA_REMOTE_NIC2 | FJMPI_RDMA_PATH0 )

Definition at line 65 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ FJMPI_RDMA_PUT_FLAGS_TO_SOUTH

#define FJMPI_RDMA_PUT_FLAGS_TO_SOUTH   (FJMPI_RDMA_LOCAL_NIC2 | FJMPI_RDMA_REMOTE_NIC0 | FJMPI_RDMA_PATH0 )

Definition at line 66 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ FJMPI_RDMA_PUT_FLAGS_TO_EAST

#define FJMPI_RDMA_PUT_FLAGS_TO_EAST   (FJMPI_RDMA_LOCAL_NIC3 | FJMPI_RDMA_REMOTE_NIC1 | FJMPI_RDMA_PATH0 )

Definition at line 67 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

◆ FJMPI_RDMA_PUT_FLAGS_TO_WEST

#define FJMPI_RDMA_PUT_FLAGS_TO_WEST   (FJMPI_RDMA_LOCAL_NIC1 | FJMPI_RDMA_REMOTE_NIC3 | FJMPI_RDMA_PATH0 )

Definition at line 68 of file scale_rdma.c.

Referenced by rdma_put8_(), and rdma_put_().

Typedef Documentation

◆ sf_t

typedef uint64_t sf_t

Definition at line 42 of file scale_rdma.c.

Function Documentation

◆ rdma_setup_()

void rdma_setup_ ( const int32_t *  COMM_vsize_max_in,
const int32_t *  IA_in,
const int32_t *  JA_in,
const int32_t *  KA_in,
const int32_t *  IHALO_in,
const int32_t *  JHALO_in,
const int32_t *  IS_in,
const int32_t *  IE_in,
const int32_t *  JS_in,
const int32_t *  JE_in,
const int32_t *  RANK_W_in,
const int32_t *  RANK_N_in,
const int32_t *  RANK_E_in,
const int32_t *  RANK_S_in 
)

Definition at line 72 of file scale_rdma.c.

References BEARING_CNT, COMM_vsize_max, datasize_NS, datasize_WE, EAST, FLAG_CNT, IA, IE, IHALO, IS, JA, JE, JHALO, JS, KA, local_sf, lvar, memid, memid_cnt, memid_sf, NORTH, RANK_E, RANK_N, RANK_S, RANK_W, rdma_put_id, remote_sf, rvar, sending, SOUTH, status_flag, mod_atmos_vars::v, and WEST.

87 {
88  int v;
89 
90  COMM_vsize_max = *COMM_vsize_max_in ;
91  IA = *IA_in ;
92  JA = *JA_in ;
93  KA = *KA_in ;
94  IHALO = *IHALO_in ;
95  JHALO = *JHALO_in ;
96  IS = *IS_in ;
97  IE = *IE_in ;
98  JS = *JS_in ;
99  JE = *JE_in ;
100  RANK_W = *RANK_W_in ;
101  RANK_N = *RANK_N_in ;
102  RANK_E = *RANK_E_in ;
103  RANK_S = *RANK_S_in ;
104 
105  datasize_NS = sizeof(var_t) * (IE-IS+1) * KA ;
106  datasize_WE = sizeof(var_t) * IHALO * KA ;
107 
108  memid_cnt = 0;
109 
110  memid = (int32_t *) malloc(sizeof(int32_t) * COMM_vsize_max);
111  lvar = (uint64_t *) malloc(sizeof(uint64_t ) * COMM_vsize_max) ;
112  rvar = (uint64_t **) malloc(sizeof(uint64_t *) * COMM_vsize_max) ;
113  for(v=0; v<COMM_vsize_max; v++) rvar[v] = (uint64_t *) malloc(sizeof(uint64_t) * BEARING_CNT) ;
114 
115  status_flag = (sf_t *) calloc(BEARING_CNT*FLAG_CNT, sizeof(sf_t)) ;
116  remote_sf = (uint64_t *) malloc(sizeof(uint64_t) * BEARING_CNT) ;
117 
118  sending = (int64_t *) calloc(BEARING_CNT, sizeof(int64_t)) ;
119 
120  rdma_put_id = 0 ;
121 
122  FJMPI_Rdma_init() ;
123 
124  memid_sf = memid_cnt ;
125  local_sf = FJMPI_Rdma_reg_mem(memid_cnt, status_flag, sizeof(sf_t)*BEARING_CNT*FLAG_CNT) ;
126  memid_cnt++ ;
127 
128  MPI_Barrier(MPI_COMM_WORLD) ;
129 
130  if( RANK_W != MPI_PROC_NULL )
131  remote_sf[WEST] = FJMPI_Rdma_get_remote_addr(RANK_W, memid_sf) ;
132  else
133  remote_sf[WEST] = 0 ;
134 
135  if( RANK_N != MPI_PROC_NULL )
136  remote_sf[NORTH] = FJMPI_Rdma_get_remote_addr(RANK_N, memid_sf) ;
137  else
138  remote_sf[NORTH] = 0 ;
139 
140  if( RANK_E != MPI_PROC_NULL )
141  remote_sf[EAST] = FJMPI_Rdma_get_remote_addr(RANK_E, memid_sf) ;
142  else
143  remote_sf[EAST] = 0 ;
144 
145  if( RANK_S != MPI_PROC_NULL )
146  remote_sf[SOUTH] = FJMPI_Rdma_get_remote_addr(RANK_S, memid_sf) ;
147  else
148  remote_sf[SOUTH] = 0 ;
149 
150 }
status_flag
volatile sf_t * status_flag
Definition: scale_rdma.c:43
RANK_N
int32_t RANK_N
Definition: scale_rdma.c:29
BEARING_CNT
#define BEARING_CNT
Definition: scale_rdma.c:18
SOUTH
#define SOUTH
Definition: scale_rdma.c:17
local_sf
uint64_t local_sf
Definition: scale_rdma.c:45
memid_sf
int32_t memid_sf
Definition: scale_rdma.c:44
RANK_S
int32_t RANK_S
Definition: scale_rdma.c:29
rdma_put_id
volatile sf_t rdma_put_id
Definition: scale_rdma.c:61
JE
int32_t JE
Definition: scale_rdma.c:24
RANK_W
int32_t RANK_W
Definition: scale_rdma.c:29
EAST
#define EAST
Definition: scale_rdma.c:16
datasize_NS
size_t datasize_NS
Definition: scale_rdma.c:27
WEST
#define WEST
Definition: scale_rdma.c:14
datasize_WE
size_t datasize_WE
Definition: scale_rdma.c:27
NORTH
#define NORTH
Definition: scale_rdma.c:15
IA
int32_t IA
Definition: scale_rdma.c:22
RANK_E
int32_t RANK_E
Definition: scale_rdma.c:29
sf_t
uint64_t sf_t
Definition: scale_rdma.c:42
JA
int32_t JA
Definition: scale_rdma.c:22
memid_cnt
int32_t memid_cnt
Definition: scale_rdma.c:35
FLAG_CNT
#define FLAG_CNT
Definition: scale_rdma.c:52
JS
int32_t JS
Definition: scale_rdma.c:24
mod_atmos_vars::v
real(rp), dimension(:,:,:), allocatable, public v
Definition: mod_atmos_vars.f90:97
IHALO
int32_t IHALO
Definition: scale_rdma.c:23
COMM_vsize_max
int32_t COMM_vsize_max
Definition: scale_rdma.c:20
remote_sf
uint64_t * remote_sf
Definition: scale_rdma.c:46
IE
int32_t IE
Definition: scale_rdma.c:24
rvar
uint64_t ** rvar
Definition: scale_rdma.c:39
JHALO
int32_t JHALO
Definition: scale_rdma.c:23
lvar
uint64_t * lvar
Definition: scale_rdma.c:38
sending
int64_t * sending
Definition: scale_rdma.c:56
IS
int32_t IS
Definition: scale_rdma.c:24
KA
int32_t KA
Definition: scale_rdma.c:22
memid
int32_t * memid
Definition: scale_rdma.c:37

◆ set_rdma_variable_()

void set_rdma_variable_ ( const var_t *  var,
const int32_t *  vid 
)

Definition at line 153 of file scale_rdma.c.

References EAST, IA, JA, KA, lvar, memid, memid_cnt, NORTH, RANK_E, RANK_N, RANK_S, RANK_W, rvar, SOUTH, and WEST.

156 {
157 
158  memid[*vid] = memid_cnt ;
159  lvar[*vid] = FJMPI_Rdma_reg_mem(memid_cnt, var, sizeof(var_t)*IA*JA*KA) ;
160  memid_cnt++ ;
161 
162 
163  MPI_Barrier(MPI_COMM_WORLD) ;
164 
165  if( RANK_W != MPI_PROC_NULL ) rvar[*vid][WEST] = FJMPI_Rdma_get_remote_addr(RANK_W, memid[*vid]) ;
166  if( RANK_N != MPI_PROC_NULL ) rvar[*vid][NORTH] = FJMPI_Rdma_get_remote_addr(RANK_N, memid[*vid]) ;
167  if( RANK_E != MPI_PROC_NULL ) rvar[*vid][EAST] = FJMPI_Rdma_get_remote_addr(RANK_E, memid[*vid]) ;
168  if( RANK_S != MPI_PROC_NULL ) rvar[*vid][SOUTH] = FJMPI_Rdma_get_remote_addr(RANK_S, memid[*vid]) ;
169 
170 }
RANK_N
int32_t RANK_N
Definition: scale_rdma.c:29
SOUTH
#define SOUTH
Definition: scale_rdma.c:17
RANK_S
int32_t RANK_S
Definition: scale_rdma.c:29
RANK_W
int32_t RANK_W
Definition: scale_rdma.c:29
EAST
#define EAST
Definition: scale_rdma.c:16
WEST
#define WEST
Definition: scale_rdma.c:14
NORTH
#define NORTH
Definition: scale_rdma.c:15
IA
int32_t IA
Definition: scale_rdma.c:22
RANK_E
int32_t RANK_E
Definition: scale_rdma.c:29
JA
int32_t JA
Definition: scale_rdma.c:22
memid_cnt
int32_t memid_cnt
Definition: scale_rdma.c:35
rvar
uint64_t ** rvar
Definition: scale_rdma.c:39
lvar
uint64_t * lvar
Definition: scale_rdma.c:38
KA
int32_t KA
Definition: scale_rdma.c:22
memid
int32_t * memid
Definition: scale_rdma.c:37

◆ rdma_put_()

void rdma_put_ ( const int32_t *  vid,
const int32_t *  num 
)

Definition at line 173 of file scale_rdma.c.

References datasize_NS, datasize_WE, EAST, FALSE, FJMPI_RDMA_PUT_FLAGS_TO_EAST, FJMPI_RDMA_PUT_FLAGS_TO_NORTH, FJMPI_RDMA_PUT_FLAGS_TO_SOUTH, FJMPI_RDMA_PUT_FLAGS_TO_WEST, IE, IHALO, IS, JE, JHALO, JS, LOCAL_PUT_DONE, LOCAL_RECV_DONE, LOCAL_RECV_READY, local_sf, lvar, NORTH, offset, RANK_E, RANK_N, RANK_S, RANK_W, rdma_put_id, RDMA_PUT_ID_MAX, RDMA_TAG, RDMA_TAG_TAIL, REMOTE_PUT_DONE, REMOTE_RECV_READY, remote_sf, rvar, sending, sidx, soffset, SOUTH, status_flag, TRUE, mod_atmos_vars::v, and WEST.

174 {
175  struct FJMPI_Rdma_cq cq ;
176  int j , v;
177 
178  rdma_put_id = (rdma_put_id % RDMA_PUT_ID_MAX) + 1 ;
179 
180  /* set status_flag(recv) */
181  status_flag[sidx(WEST ,LOCAL_RECV_READY)] = rdma_put_id ;
182  status_flag[sidx(NORTH,LOCAL_RECV_READY)] = rdma_put_id ;
183  status_flag[sidx(EAST ,LOCAL_RECV_READY)] = rdma_put_id ;
184  status_flag[sidx(SOUTH,LOCAL_RECV_READY)] = rdma_put_id ;
185 
186  if( RANK_S != MPI_PROC_NULL )
187  FJMPI_Rdma_put(RANK_S, RDMA_TAG,
188  remote_sf[SOUTH]+soffset(NORTH,REMOTE_RECV_READY),
189  local_sf+soffset(SOUTH,LOCAL_RECV_READY),
190  sizeof(sf_t),
191  FJMPI_RDMA_PUT_FLAGS_TO_SOUTH | FJMPI_RDMA_STRONG_ORDER ) ;
192 
193  if( RANK_N != MPI_PROC_NULL )
194  FJMPI_Rdma_put(RANK_N, RDMA_TAG,
195  remote_sf[NORTH]+soffset(SOUTH,REMOTE_RECV_READY),
196  local_sf+soffset(NORTH,LOCAL_RECV_READY),
197  sizeof(sf_t),
198  FJMPI_RDMA_PUT_FLAGS_TO_NORTH | FJMPI_RDMA_STRONG_ORDER ) ;
199 
200  if( RANK_E != MPI_PROC_NULL )
201  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
202  remote_sf[EAST]+soffset(WEST,REMOTE_RECV_READY),
203  local_sf+soffset(EAST,LOCAL_RECV_READY),
204  sizeof(sf_t),
205  FJMPI_RDMA_PUT_FLAGS_TO_EAST | FJMPI_RDMA_STRONG_ORDER ) ;
206 
207  if( RANK_W != MPI_PROC_NULL )
208  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
209  remote_sf[WEST]+soffset(EAST,REMOTE_RECV_READY),
210  local_sf+soffset(WEST,LOCAL_RECV_READY),
211  sizeof(sf_t),
212  FJMPI_RDMA_PUT_FLAGS_TO_WEST | FJMPI_RDMA_STRONG_ORDER ) ;
213 
214  /* send data */
215  do {
216  // to north
217  if( status_flag[sidx(NORTH,LOCAL_PUT_DONE)] != rdma_put_id && RANK_N == MPI_PROC_NULL )
218  {
219  status_flag[sidx(NORTH,LOCAL_PUT_DONE)] = rdma_put_id ;
220  status_flag[sidx(NORTH,LOCAL_RECV_DONE)] = rdma_put_id ;
221  sending[NORTH] = FALSE ;
222  }
223 
224  if( status_flag[sidx(NORTH,LOCAL_PUT_DONE)] != rdma_put_id &&
225  status_flag[sidx(NORTH,REMOTE_RECV_READY)] == rdma_put_id )
226  {
227  sending[NORTH] = TRUE ;
228 
229  // put_data
230  for(v=0; v<*num; v++)
231  {
232  for(j=0; j<JHALO; j++)
233  {
234  FJMPI_Rdma_put(RANK_N, RDMA_TAG,
235  rvar[*vid+v][NORTH]+offset(0,IS-1,j+JE),
236  lvar[*vid+v]+offset(0,IS-1,j+JS-1),
237  datasize_NS,
238  FJMPI_RDMA_PUT_FLAGS_TO_NORTH ) ;
239  }
240  }
241 
242  // set status_flag(send)
243  status_flag[sidx(NORTH,LOCAL_PUT_DONE)] = rdma_put_id ;
244  FJMPI_Rdma_put(RANK_N, RDMA_TAG,
245  remote_sf[NORTH]+soffset(SOUTH,REMOTE_PUT_DONE),
246  local_sf+soffset(NORTH,LOCAL_PUT_DONE),
247  sizeof(sf_t),
248  FJMPI_RDMA_PUT_FLAGS_TO_NORTH | FJMPI_RDMA_STRONG_ORDER ) ;
249  }
250 
251  // to south
252  if( status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] != rdma_put_id && RANK_S == MPI_PROC_NULL )
253  {
254  status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] = rdma_put_id ;
255  status_flag[sidx(SOUTH,LOCAL_RECV_DONE)] = rdma_put_id ;
256  sending[SOUTH] = FALSE;
257  }
258 
259  if( status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] != rdma_put_id &&
260  status_flag[sidx(SOUTH,REMOTE_RECV_READY)] == rdma_put_id )
261  {
262  sending[SOUTH] = TRUE;
263 
264  // put_data
265  for(v=0; v<*num; v++)
266  {
267  for(j=0; j<JHALO; j++)
268  {
269  FJMPI_Rdma_put(RANK_S, RDMA_TAG,
270  rvar[*vid+v][SOUTH]+offset(0,IS-1,j+JS-JHALO-1),
271  lvar[*vid+v]+offset(0,IS-1,j+JE-JHALO),
272  datasize_NS,
273  FJMPI_RDMA_PUT_FLAGS_TO_SOUTH ) ;
274  }
275  }
276 
277  // set status_flag(send)
278  status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] = rdma_put_id ;
279  FJMPI_Rdma_put(RANK_S, RDMA_TAG,
280  remote_sf[SOUTH]+soffset(NORTH,REMOTE_PUT_DONE),
281  local_sf+soffset(SOUTH,LOCAL_PUT_DONE),
282  sizeof(sf_t),
283  FJMPI_RDMA_PUT_FLAGS_TO_SOUTH | FJMPI_RDMA_STRONG_ORDER ) ;
284  }
285 
286  // to west
287  if( status_flag[sidx(WEST,LOCAL_PUT_DONE)] != rdma_put_id && RANK_W == MPI_PROC_NULL )
288  {
289  status_flag[sidx(WEST,LOCAL_PUT_DONE)] = rdma_put_id ;
290  status_flag[sidx(WEST,LOCAL_RECV_DONE)] = rdma_put_id ;
291  sending[WEST] = FALSE;
292  }
293 
294  if( status_flag[sidx(WEST,LOCAL_PUT_DONE)] != rdma_put_id &&
295  status_flag[sidx(WEST,REMOTE_RECV_READY)] == rdma_put_id )
296  {
297  sending[WEST] = TRUE;
298 
299  // put_data
300  for(v=0; v<*num; v++)
301  {
302  for(j=JS-1; j<JE; j++)
303  {
304  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
305  rvar[*vid+v][WEST]+offset(0,IE,j),
306  lvar[*vid+v]+offset(0,IS-1,j),
307  datasize_WE,
308  FJMPI_RDMA_PUT_FLAGS_TO_WEST ) ;
309  }
310  }
311 
312  // set status_flag(send)
313  status_flag[sidx(WEST,LOCAL_PUT_DONE)] = rdma_put_id ;
314  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
315  remote_sf[WEST]+soffset(EAST,REMOTE_PUT_DONE),
316  local_sf+soffset(WEST,LOCAL_PUT_DONE),
317  sizeof(sf_t),
318  FJMPI_RDMA_PUT_FLAGS_TO_WEST | FJMPI_RDMA_STRONG_ORDER ) ;
319  }
320 
321  // to east
322  if( status_flag[sidx(EAST,LOCAL_PUT_DONE)] != rdma_put_id && RANK_E == MPI_PROC_NULL )
323  {
324  status_flag[sidx(EAST,LOCAL_PUT_DONE)] = rdma_put_id ;
325  status_flag[sidx(EAST,LOCAL_RECV_DONE)] = rdma_put_id ;
326  sending[EAST] = FALSE;
327  }
328 
329  if( status_flag[sidx(EAST,LOCAL_PUT_DONE)] != rdma_put_id &&
330  status_flag[sidx(EAST,REMOTE_RECV_READY)] == rdma_put_id )
331  {
332  sending[TRUE] = FALSE;
333 
334  // put_data
335  for(v=0; v<*num; v++)
336  {
337  for(j=JS-1; j<JE; j++)
338  {
339  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
340  rvar[*vid+v][EAST]+offset(0,IS-IHALO-1,j),
341  lvar[*vid+v]+offset(0,IE-IHALO,j),
342  datasize_WE,
343  FJMPI_RDMA_PUT_FLAGS_TO_EAST ) ;
344  }
345  }
346 
347  // set status_flag(send)
348  status_flag[sidx(EAST,LOCAL_PUT_DONE)] = rdma_put_id ;
349  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
350  remote_sf[EAST]+soffset(WEST,REMOTE_PUT_DONE),
351  local_sf+soffset(EAST,LOCAL_PUT_DONE),
352  sizeof(sf_t),
353  FJMPI_RDMA_PUT_FLAGS_TO_EAST | FJMPI_RDMA_STRONG_ORDER ) ;
354  }
355  } while( status_flag[sidx(WEST, LOCAL_PUT_DONE)] != rdma_put_id ||
356  status_flag[sidx(NORTH,LOCAL_PUT_DONE)] != rdma_put_id ||
357  status_flag[sidx(EAST, LOCAL_PUT_DONE)] != rdma_put_id ||
358  status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] != rdma_put_id ) ;
359 
360  /* completion check (put) */
361  do {
362  // to west
363  if( RANK_W != MPI_PROC_NULL ) {
364  while(FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC1, &cq) == FJMPI_RDMA_NOTICE ) {
365  if(cq.tag == RDMA_TAG_TAIL) sending[WEST] = FALSE ;
366  }
367  }
368  // to east
369  if( RANK_E != MPI_PROC_NULL ) {
370  while(FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC3, &cq) == FJMPI_RDMA_NOTICE ) {
371  if(cq.tag == RDMA_TAG_TAIL) sending[EAST] = FALSE ;
372  }
373  }
374  // to north
375  if( RANK_N != MPI_PROC_NULL ) {
376  while(FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC0, &cq) == FJMPI_RDMA_NOTICE ) {
377  if(cq.tag == RDMA_TAG_TAIL) sending[NORTH] = FALSE ;
378  }
379  }
380  // to south
381  if( RANK_S != MPI_PROC_NULL ) {
382  while(FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC2, &cq) == FJMPI_RDMA_NOTICE ) {
383  if(cq.tag == RDMA_TAG_TAIL) sending[SOUTH] = FALSE ;
384  }
385  }
386  } while(sending[WEST] || sending[NORTH] || sending[EAST] || sending[SOUTH] ) ;
387 
388  /* completion check (recv) */
389  while( status_flag[sidx(WEST,LOCAL_RECV_DONE)] != rdma_put_id ||
390  status_flag[sidx(NORTH,LOCAL_RECV_DONE)] != rdma_put_id ||
391  status_flag[sidx(EAST,LOCAL_RECV_DONE)] != rdma_put_id ||
392  status_flag[sidx(SOUTH,LOCAL_RECV_DONE)] != rdma_put_id ) ;
393 
394 }
status_flag
volatile sf_t * status_flag
Definition: scale_rdma.c:43
RANK_N
int32_t RANK_N
Definition: scale_rdma.c:29
FALSE
#define FALSE
Definition: scale_rdma.c:57
LOCAL_RECV_READY
#define LOCAL_RECV_READY
Definition: scale_rdma.c:47
REMOTE_RECV_READY
#define REMOTE_RECV_READY
Definition: scale_rdma.c:48
SOUTH
#define SOUTH
Definition: scale_rdma.c:17
LOCAL_RECV_DONE
#define LOCAL_RECV_DONE
Definition: scale_rdma.c:51
local_sf
uint64_t local_sf
Definition: scale_rdma.c:45
RANK_S
int32_t RANK_S
Definition: scale_rdma.c:29
rdma_put_id
volatile sf_t rdma_put_id
Definition: scale_rdma.c:61
JE
int32_t JE
Definition: scale_rdma.c:24
RANK_W
int32_t RANK_W
Definition: scale_rdma.c:29
EAST
#define EAST
Definition: scale_rdma.c:16
soffset
#define soffset(DIR, FLAG)
Definition: scale_rdma.c:54
REMOTE_PUT_DONE
#define REMOTE_PUT_DONE
Definition: scale_rdma.c:50
datasize_NS
size_t datasize_NS
Definition: scale_rdma.c:27
TRUE
#define TRUE
Definition: scale_rdma.c:58
WEST
#define WEST
Definition: scale_rdma.c:14
datasize_WE
size_t datasize_WE
Definition: scale_rdma.c:27
NORTH
#define NORTH
Definition: scale_rdma.c:15
RANK_E
int32_t RANK_E
Definition: scale_rdma.c:29
sf_t
uint64_t sf_t
Definition: scale_rdma.c:42
RDMA_TAG
#define RDMA_TAG
Definition: scale_rdma.c:32
LOCAL_PUT_DONE
#define LOCAL_PUT_DONE
Definition: scale_rdma.c:49
JS
int32_t JS
Definition: scale_rdma.c:24
FJMPI_RDMA_PUT_FLAGS_TO_NORTH
#define FJMPI_RDMA_PUT_FLAGS_TO_NORTH
Definition: scale_rdma.c:65
mod_atmos_vars::v
real(rp), dimension(:,:,:), allocatable, public v
Definition: mod_atmos_vars.f90:97
FJMPI_RDMA_PUT_FLAGS_TO_SOUTH
#define FJMPI_RDMA_PUT_FLAGS_TO_SOUTH
Definition: scale_rdma.c:66
RDMA_PUT_ID_MAX
#define RDMA_PUT_ID_MAX
Definition: scale_rdma.c:62
offset
#define offset(Z, X, Y)
Definition: scale_rdma.c:25
IHALO
int32_t IHALO
Definition: scale_rdma.c:23
remote_sf
uint64_t * remote_sf
Definition: scale_rdma.c:46
IE
int32_t IE
Definition: scale_rdma.c:24
rvar
uint64_t ** rvar
Definition: scale_rdma.c:39
JHALO
int32_t JHALO
Definition: scale_rdma.c:23
sidx
#define sidx(DIR, FLAG)
Definition: scale_rdma.c:53
lvar
uint64_t * lvar
Definition: scale_rdma.c:38
FJMPI_RDMA_PUT_FLAGS_TO_WEST
#define FJMPI_RDMA_PUT_FLAGS_TO_WEST
Definition: scale_rdma.c:68
RDMA_TAG_TAIL
#define RDMA_TAG_TAIL
Definition: scale_rdma.c:33
sending
int64_t * sending
Definition: scale_rdma.c:56
IS
int32_t IS
Definition: scale_rdma.c:24
FJMPI_RDMA_PUT_FLAGS_TO_EAST
#define FJMPI_RDMA_PUT_FLAGS_TO_EAST
Definition: scale_rdma.c:67

◆ rdma_put8_()

void rdma_put8_ ( const int32_t *  vid,
const int32_t *  num 
)

Definition at line 396 of file scale_rdma.c.

References datasize_NS, datasize_WE, EAST, FALSE, FJMPI_RDMA_PUT_FLAGS_TO_EAST, FJMPI_RDMA_PUT_FLAGS_TO_NORTH, FJMPI_RDMA_PUT_FLAGS_TO_SOUTH, FJMPI_RDMA_PUT_FLAGS_TO_WEST, IE, IHALO, IS, JE, JHALO, JS, LOCAL_PUT_DONE, LOCAL_RECV_DONE, LOCAL_RECV_READY, local_sf, lvar, NORTH, offset, RANK_E, RANK_N, RANK_S, RANK_W, rdma_put_id, RDMA_PUT_ID_MAX, RDMA_TAG, RDMA_TAG_TAIL, REMOTE_PUT_DONE, REMOTE_RECV_READY, remote_sf, rvar, sending, sidx, soffset, SOUTH, status_flag, TRUE, mod_atmos_vars::v, and WEST.

397 {
398  struct FJMPI_Rdma_cq cq ;
399  int j , v;
400 
401  rdma_put_id = (rdma_put_id % RDMA_PUT_ID_MAX) + 1 ;
402 
403  /* set status_flag(recv) */
404  status_flag[sidx(WEST ,LOCAL_RECV_READY)] = rdma_put_id ;
405  status_flag[sidx(NORTH,LOCAL_RECV_READY)] = rdma_put_id ;
406  status_flag[sidx(EAST ,LOCAL_RECV_READY)] = rdma_put_id ;
407  status_flag[sidx(SOUTH,LOCAL_RECV_READY)] = rdma_put_id ;
408 
409  if( RANK_S != MPI_PROC_NULL )
410  FJMPI_Rdma_put(RANK_S, RDMA_TAG,
411  remote_sf[SOUTH]+soffset(NORTH,REMOTE_RECV_READY),
412  local_sf+soffset(SOUTH,LOCAL_RECV_READY),
413  sizeof(sf_t),
414  FJMPI_RDMA_PUT_FLAGS_TO_SOUTH | FJMPI_RDMA_STRONG_ORDER ) ;
415 
416  if( RANK_N != MPI_PROC_NULL )
417  FJMPI_Rdma_put(RANK_N, RDMA_TAG,
418  remote_sf[NORTH]+soffset(SOUTH,REMOTE_RECV_READY),
419  local_sf+soffset(NORTH,LOCAL_RECV_READY),
420  sizeof(sf_t),
421  FJMPI_RDMA_PUT_FLAGS_TO_NORTH | FJMPI_RDMA_STRONG_ORDER ) ;
422 
423  if( RANK_E != MPI_PROC_NULL )
424  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
425  remote_sf[EAST]+soffset(WEST,REMOTE_RECV_READY),
426  local_sf+soffset(EAST,LOCAL_RECV_READY),
427  sizeof(sf_t),
428  FJMPI_RDMA_PUT_FLAGS_TO_EAST | FJMPI_RDMA_STRONG_ORDER ) ;
429 
430  if( RANK_W != MPI_PROC_NULL )
431  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
432  remote_sf[WEST]+soffset(EAST,REMOTE_RECV_READY),
433  local_sf+soffset(WEST,LOCAL_RECV_READY),
434  sizeof(sf_t),
435  FJMPI_RDMA_PUT_FLAGS_TO_WEST | FJMPI_RDMA_STRONG_ORDER ) ;
436 
437  /* send data */
438  do {
439  // to north
440  if( status_flag[sidx(NORTH,LOCAL_PUT_DONE)] != rdma_put_id && RANK_N == MPI_PROC_NULL )
441  {
442  status_flag[sidx(NORTH,LOCAL_PUT_DONE)] = rdma_put_id ;
443  sending[NORTH] = FALSE ;
444  }
445 
446  if( status_flag[sidx(NORTH,LOCAL_PUT_DONE)] != rdma_put_id &&
447  status_flag[sidx(NORTH,REMOTE_RECV_READY)] == rdma_put_id )
448  {
449  sending[NORTH] = TRUE ;
450 
451  // put_data
452  for(v=0; v<*num; v++)
453  {
454  for(j=0; j<JHALO; j++)
455  {
456  FJMPI_Rdma_put(RANK_N, RDMA_TAG,
457  rvar[*vid+v][NORTH]+offset(0,IS-1,j+JE),
458  lvar[*vid+v]+offset(0,IS-1,j+JS-1),
459  datasize_NS,
460  FJMPI_RDMA_PUT_FLAGS_TO_NORTH ) ;
461 
462  }
463  }
464 
465  // set status_flag(send)
466  status_flag[sidx(NORTH,LOCAL_PUT_DONE)] = rdma_put_id ;
467  FJMPI_Rdma_put(RANK_N, RDMA_TAG_TAIL,
468  remote_sf[NORTH]+soffset(SOUTH,REMOTE_PUT_DONE),
469  local_sf+soffset(NORTH,LOCAL_PUT_DONE),
470  sizeof(sf_t),
471  FJMPI_RDMA_PUT_FLAGS_TO_NORTH | FJMPI_RDMA_STRONG_ORDER ) ;
472  }
473 
474  // to south
475  if( status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] != rdma_put_id && RANK_S == MPI_PROC_NULL )
476  {
477  status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] = rdma_put_id ;
478  sending[SOUTH] = FALSE;
479  }
480 
481  if( status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] != rdma_put_id &&
482  status_flag[sidx(SOUTH,REMOTE_RECV_READY)] == rdma_put_id )
483  {
484  sending[SOUTH] = TRUE;
485 
486  // put_data
487  for(v=0; v<*num; v++)
488  {
489  for(j=0; j<JHALO; j++)
490  {
491  FJMPI_Rdma_put(RANK_S, RDMA_TAG,
492  rvar[*vid+v][SOUTH]+offset(0,IS-1,j+JS-JHALO-1),
493  lvar[*vid+v]+offset(0,IS-1,j+JE-JHALO),
494  datasize_NS,
495  FJMPI_RDMA_PUT_FLAGS_TO_SOUTH ) ;
496  }
497  }
498 
499  // set status_flag(send)
500  status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] = rdma_put_id ;
501  FJMPI_Rdma_put(RANK_S, RDMA_TAG_TAIL,
502  remote_sf[SOUTH]+soffset(NORTH,REMOTE_PUT_DONE),
503  local_sf+soffset(SOUTH,LOCAL_PUT_DONE),
504  sizeof(sf_t),
505  FJMPI_RDMA_PUT_FLAGS_TO_SOUTH | FJMPI_RDMA_STRONG_ORDER ) ;
506  }
507 
508  // to west
509  if( status_flag[sidx(WEST,LOCAL_PUT_DONE)] != rdma_put_id && RANK_W == MPI_PROC_NULL )
510  {
511  status_flag[sidx(WEST,LOCAL_PUT_DONE)] = rdma_put_id ;
512  sending[WEST] = FALSE ;
513  }
514 
515  if( status_flag[sidx(WEST,LOCAL_PUT_DONE)] != rdma_put_id &&
516  status_flag[sidx(WEST,REMOTE_RECV_READY)] == rdma_put_id )
517  {
518  sending[WEST] = TRUE ;
519 
520  // put_data
521  for(v=0; v<*num; v++)
522  {
523  for(j=JS-1; j<JE; j++)
524  {
525  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
526  rvar[*vid+v][WEST]+offset(0,IE,j),
527  lvar[*vid+v]+offset(0,IS-1,j),
528  datasize_WE,
529  FJMPI_RDMA_PUT_FLAGS_TO_WEST ) ;
530  }
531  }
532 
533  // set status_flag(send)
534  status_flag[sidx(WEST,LOCAL_PUT_DONE)] = rdma_put_id ;
535  }
536 
537  // to east
538  if( status_flag[sidx(EAST,LOCAL_PUT_DONE)] != rdma_put_id && RANK_E == MPI_PROC_NULL )
539  {
540  status_flag[sidx(EAST,LOCAL_PUT_DONE)] = rdma_put_id ;
541  sending[EAST] = FALSE ;
542  }
543 
544  if( status_flag[sidx(EAST,LOCAL_PUT_DONE)] != rdma_put_id &&
545  status_flag[sidx(EAST,REMOTE_RECV_READY)] == rdma_put_id )
546  {
547  sending[EAST] = TRUE ;
548 
549  // put_data
550  for(v=0; v<*num; v++)
551  {
552  for(j=JS-1; j<JE; j++)
553  {
554  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
555  rvar[*vid+v][EAST]+offset(0,IS-IHALO-1,j),
556  lvar[*vid+v]+offset(0,IE-IHALO,j),
557  datasize_WE,
558  FJMPI_RDMA_PUT_FLAGS_TO_EAST ) ;
559  }
560  }
561 
562  // set status_flag(send)
563  status_flag[sidx(EAST,LOCAL_PUT_DONE)] = rdma_put_id ;
564  }
565 
566  } while( status_flag[sidx(WEST,LOCAL_PUT_DONE)] != rdma_put_id ||
567  status_flag[sidx(NORTH,LOCAL_PUT_DONE)] != rdma_put_id ||
568  status_flag[sidx(EAST,LOCAL_PUT_DONE)] != rdma_put_id ||
569  status_flag[sidx(SOUTH,LOCAL_PUT_DONE)] != rdma_put_id ) ;
570 
571  /* completion check (recv from NORTH and SOUTH) */
572  if( RANK_N == MPI_PROC_NULL ) status_flag[sidx(NORTH, LOCAL_RECV_DONE)] = rdma_put_id ;
573  if( RANK_S == MPI_PROC_NULL ) status_flag[sidx(SOUTH, LOCAL_RECV_DONE)] = rdma_put_id ;
574 
575  while( status_flag[sidx(NORTH,LOCAL_RECV_DONE)] != rdma_put_id ||
576  status_flag[sidx(SOUTH,LOCAL_RECV_DONE)] != rdma_put_id ) ;
577 
578  /* send data from North and South to West and East */
579  for(v=0; v<*num; v++)
580  {
581  // data form north
582  if( RANK_N != MPI_PROC_NULL ) {
583  if( RANK_W != MPI_PROC_NULL && RANK_E != MPI_PROC_NULL ) {
584  for(j=0; j<JHALO; j++)
585  {
586  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
587  rvar[*vid+v][WEST]+offset(0,IE,j),
588  lvar[*vid+v]+offset(0,IS-1,j),
589  datasize_WE,
590  FJMPI_RDMA_PUT_FLAGS_TO_WEST ) ;
591  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
592  rvar[*vid+v][EAST]+offset(0,IS-IHALO-1,j),
593  lvar[*vid+v]+offset(0,IE-IHALO,j),
594  datasize_WE,
595  FJMPI_RDMA_PUT_FLAGS_TO_EAST ) ;
596  }
597  }
598  else if( RANK_W != MPI_PROC_NULL ) {
599  for(j=0; j<JHALO; j++)
600  {
601  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
602  rvar[*vid+v][WEST]+offset(0,IE,j),
603  lvar[*vid+v]+offset(0,IS-1,j),
604  datasize_WE,
605  FJMPI_RDMA_PUT_FLAGS_TO_WEST ) ;
606  }
607  }
608  else if( RANK_E != MPI_PROC_NULL ) {
609  for(j=0; j<JHALO; j++)
610  {
611  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
612  rvar[*vid+v][EAST]+offset(0,IS-IHALO-1,j),
613  lvar[*vid+v]+offset(0,IE-IHALO,j),
614  datasize_WE,
615  FJMPI_RDMA_PUT_FLAGS_TO_EAST ) ;
616  }
617  }
618  }
619 
620  // data from south
621  if( RANK_S != MPI_PROC_NULL ) {
622  if( RANK_W != MPI_PROC_NULL && RANK_E != MPI_PROC_NULL ) {
623  for(j=JE; j<JE+JHALO; j++)
624  {
625  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
626  rvar[*vid+v][WEST]+offset(0,IE,j),
627  lvar[*vid+v]+offset(0,IS-1,j),
628  datasize_WE,
629  FJMPI_RDMA_PUT_FLAGS_TO_WEST ) ;
630  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
631  rvar[*vid+v][EAST]+offset(0,IS-IHALO-1,j),
632  lvar[*vid+v]+offset(0,IE-IHALO,j),
633  datasize_WE,
634  FJMPI_RDMA_PUT_FLAGS_TO_EAST ) ;
635  }
636  }
637  else if( RANK_W != MPI_PROC_NULL ) {
638  for(j=JE; j<JE+JHALO; j++)
639  {
640  FJMPI_Rdma_put(RANK_W, RDMA_TAG,
641  rvar[*vid+v][WEST]+offset(0,IE,j),
642  lvar[*vid+v]+offset(0,IS-1,j),
643  datasize_WE,
644  FJMPI_RDMA_PUT_FLAGS_TO_WEST ) ;
645  }
646  }
647  else if( RANK_E != MPI_PROC_NULL ) {
648  for(j=JE; j<JE+JHALO; j++)
649  {
650  FJMPI_Rdma_put(RANK_E, RDMA_TAG,
651  rvar[*vid+v][EAST]+offset(0,IS-IHALO-1,j),
652  lvar[*vid+v]+offset(0,IE-IHALO,j),
653  datasize_WE,
654  FJMPI_RDMA_PUT_FLAGS_TO_EAST ) ;
655  }
656  }
657  }
658  }
659 
660  if( RANK_W != MPI_PROC_NULL ) {
661  FJMPI_Rdma_put(RANK_W, RDMA_TAG_TAIL,
662  remote_sf[WEST]+soffset(EAST,REMOTE_PUT_DONE),
663  local_sf+soffset(WEST,LOCAL_PUT_DONE),
664  sizeof(sf_t),
665  FJMPI_RDMA_PUT_FLAGS_TO_WEST | FJMPI_RDMA_STRONG_ORDER ) ;
666  }
667 
668  if( RANK_E != MPI_PROC_NULL ) {
669  FJMPI_Rdma_put(RANK_E, RDMA_TAG_TAIL,
670  remote_sf[EAST]+soffset(WEST,REMOTE_PUT_DONE),
671  local_sf+soffset(EAST,LOCAL_PUT_DONE),
672  sizeof(sf_t),
673  FJMPI_RDMA_PUT_FLAGS_TO_EAST | FJMPI_RDMA_STRONG_ORDER ) ;
674  }
675 
676  /* completion check (put) */
677  do {
678  // to west
679  if( RANK_W != MPI_PROC_NULL ) {
680  while(FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC1, &cq) == FJMPI_RDMA_NOTICE ) {
681  if(cq.tag == RDMA_TAG_TAIL) sending[WEST] = FALSE ;
682  }
683  }
684  // to east
685  if( RANK_E != MPI_PROC_NULL ) {
686  while(FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC3, &cq) == FJMPI_RDMA_NOTICE ) {
687  if(cq.tag == RDMA_TAG_TAIL) sending[EAST] = FALSE ;
688  }
689  }
690  // to north
691  if( RANK_N != MPI_PROC_NULL ) {
692  while(FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC0, &cq) == FJMPI_RDMA_NOTICE ) {
693  if(cq.tag == RDMA_TAG_TAIL) sending[NORTH] = FALSE ;
694  }
695  }
696  // to south
697  if( RANK_S != MPI_PROC_NULL ) {
698  while(FJMPI_Rdma_poll_cq(FJMPI_RDMA_NIC2, &cq) == FJMPI_RDMA_NOTICE ) {
699  if(cq.tag == RDMA_TAG_TAIL) sending[SOUTH] = FALSE ;
700  }
701  }
702  } while(sending[WEST] || sending[NORTH] || sending[EAST] || sending[SOUTH] ) ;
703 
704  /* completion check (recv from WEST and EAST) */
705  if( RANK_W == MPI_PROC_NULL ) status_flag[sidx(WEST, LOCAL_RECV_DONE)] = rdma_put_id ;
706  if( RANK_E == MPI_PROC_NULL ) status_flag[sidx(EAST, LOCAL_RECV_DONE)] = rdma_put_id ;
707 
708 
709  while( status_flag[sidx(WEST,LOCAL_RECV_DONE)] != rdma_put_id ||
710  status_flag[sidx(EAST,LOCAL_RECV_DONE)] != rdma_put_id ) ;
711 }
status_flag
volatile sf_t * status_flag
Definition: scale_rdma.c:43
RANK_N
int32_t RANK_N
Definition: scale_rdma.c:29
FALSE
#define FALSE
Definition: scale_rdma.c:57
LOCAL_RECV_READY
#define LOCAL_RECV_READY
Definition: scale_rdma.c:47
REMOTE_RECV_READY
#define REMOTE_RECV_READY
Definition: scale_rdma.c:48
SOUTH
#define SOUTH
Definition: scale_rdma.c:17
LOCAL_RECV_DONE
#define LOCAL_RECV_DONE
Definition: scale_rdma.c:51
local_sf
uint64_t local_sf
Definition: scale_rdma.c:45
RANK_S
int32_t RANK_S
Definition: scale_rdma.c:29
rdma_put_id
volatile sf_t rdma_put_id
Definition: scale_rdma.c:61
JE
int32_t JE
Definition: scale_rdma.c:24
RANK_W
int32_t RANK_W
Definition: scale_rdma.c:29
EAST
#define EAST
Definition: scale_rdma.c:16
soffset
#define soffset(DIR, FLAG)
Definition: scale_rdma.c:54
REMOTE_PUT_DONE
#define REMOTE_PUT_DONE
Definition: scale_rdma.c:50
datasize_NS
size_t datasize_NS
Definition: scale_rdma.c:27
TRUE
#define TRUE
Definition: scale_rdma.c:58
WEST
#define WEST
Definition: scale_rdma.c:14
datasize_WE
size_t datasize_WE
Definition: scale_rdma.c:27
NORTH
#define NORTH
Definition: scale_rdma.c:15
RANK_E
int32_t RANK_E
Definition: scale_rdma.c:29
sf_t
uint64_t sf_t
Definition: scale_rdma.c:42
RDMA_TAG
#define RDMA_TAG
Definition: scale_rdma.c:32
LOCAL_PUT_DONE
#define LOCAL_PUT_DONE
Definition: scale_rdma.c:49
JS
int32_t JS
Definition: scale_rdma.c:24
FJMPI_RDMA_PUT_FLAGS_TO_NORTH
#define FJMPI_RDMA_PUT_FLAGS_TO_NORTH
Definition: scale_rdma.c:65
mod_atmos_vars::v
real(rp), dimension(:,:,:), allocatable, public v
Definition: mod_atmos_vars.f90:97
FJMPI_RDMA_PUT_FLAGS_TO_SOUTH
#define FJMPI_RDMA_PUT_FLAGS_TO_SOUTH
Definition: scale_rdma.c:66
RDMA_PUT_ID_MAX
#define RDMA_PUT_ID_MAX
Definition: scale_rdma.c:62
offset
#define offset(Z, X, Y)
Definition: scale_rdma.c:25
IHALO
int32_t IHALO
Definition: scale_rdma.c:23
remote_sf
uint64_t * remote_sf
Definition: scale_rdma.c:46
IE
int32_t IE
Definition: scale_rdma.c:24
rvar
uint64_t ** rvar
Definition: scale_rdma.c:39
JHALO
int32_t JHALO
Definition: scale_rdma.c:23
sidx
#define sidx(DIR, FLAG)
Definition: scale_rdma.c:53
lvar
uint64_t * lvar
Definition: scale_rdma.c:38
FJMPI_RDMA_PUT_FLAGS_TO_WEST
#define FJMPI_RDMA_PUT_FLAGS_TO_WEST
Definition: scale_rdma.c:68
RDMA_TAG_TAIL
#define RDMA_TAG_TAIL
Definition: scale_rdma.c:33
sending
int64_t * sending
Definition: scale_rdma.c:56
IS
int32_t IS
Definition: scale_rdma.c:24
FJMPI_RDMA_PUT_FLAGS_TO_EAST
#define FJMPI_RDMA_PUT_FLAGS_TO_EAST
Definition: scale_rdma.c:67

Variable Documentation

◆ COMM_vsize_max

int32_t COMM_vsize_max

Definition at line 20 of file scale_rdma.c.

Referenced by rdma_setup_().

◆ IA

int32_t IA

Definition at line 22 of file scale_rdma.c.

Referenced by rdma_setup_(), and set_rdma_variable_().

◆ JA

int32_t JA

Definition at line 22 of file scale_rdma.c.

Referenced by rdma_setup_(), and set_rdma_variable_().

◆ KA

int32_t KA

Definition at line 22 of file scale_rdma.c.

Referenced by rdma_setup_(), and set_rdma_variable_().

◆ IHALO

int32_t IHALO

Definition at line 23 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ JHALO

int32_t JHALO

Definition at line 23 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ IS

int32_t IS

Definition at line 24 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ IE

int32_t IE

Definition at line 24 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ JS

int32_t JS

Definition at line 24 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ JE

int32_t JE

Definition at line 24 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ datasize_NS

size_t datasize_NS

Definition at line 27 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ datasize_WE

size_t datasize_WE

Definition at line 27 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ RANK_W

int32_t RANK_W

Definition at line 29 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ RANK_N

int32_t RANK_N

Definition at line 29 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ RANK_E

int32_t RANK_E

Definition at line 29 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ RANK_S

int32_t RANK_S

Definition at line 29 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ memid_cnt

int32_t memid_cnt

Definition at line 35 of file scale_rdma.c.

Referenced by rdma_setup_(), and set_rdma_variable_().

◆ memid

int32_t* memid

Definition at line 37 of file scale_rdma.c.

Referenced by rdma_setup_(), and set_rdma_variable_().

◆ lvar

uint64_t* lvar

Definition at line 38 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ rvar

uint64_t** rvar

Definition at line 39 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), rdma_setup_(), and set_rdma_variable_().

◆ status_flag

volatile sf_t* status_flag

Definition at line 43 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ memid_sf

int32_t memid_sf

Definition at line 44 of file scale_rdma.c.

Referenced by rdma_setup_().

◆ local_sf

uint64_t local_sf

Definition at line 45 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ remote_sf

uint64_t* remote_sf

Definition at line 46 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ sending

int64_t* sending

Definition at line 56 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

◆ rdma_put_id

volatile sf_t rdma_put_id

Definition at line 61 of file scale_rdma.c.

Referenced by rdma_put8_(), rdma_put_(), and rdma_setup_().

Generated by   doxygen 1.8.13