1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
|
#include <stdio.h>
#include "MPIutils.hh"
#include "MPIO.hh"
void MPIO::requestSlice(int z,float *slicebuffer,MPI_Request *req){
// skip to z
//loop (foreach p of pe's)
if(!isRoot()) return;
//printf("requestSlice %u",z);
for(int reqindex=0,offset=0,idx=slicemap[z],lastchunk=slicemap[z]+chunkindexcount[z];
idx<lastchunk;idx++,reqindex++){
int p = procmap[idx];
int *idims = gdims+chunkindex[idx];
int *iorigin = gorigins + chunkindex[idx];
comm.iRecv(p,z,idims[0]*idims[1],slicebuffer+offset,req[reqindex]);
//printf("reqslice(%u):MPI_Request=%u\n",z,req[reqindex]);
offset+=idims[0]*idims[1];
}
// break if outside of z-range (dangerous if non-canonical proc layout)
}
void MPIO::waitForSlice(int z,float *slicebuffer,float *destbuffer,MPI_Request *req){
if(!isRoot()) return;
// printf("waitforslice %u\n",z);
// could do this 8k at a time
// loop for each p of pe's
// if within z-range, compute nelem based on dims at p
// wait for current request.
// re-copy based on offset & origin
sliceallwait.start();
//printf("z(%u) slicemap(%u) len(%u)\n",z,slicemap[z],chunkindexcount[z]);
for(int reqindex=0,chunkidx=slicemap[z],lastchunk=slicemap[z]+chunkindexcount[z];
chunkidx<lastchunk;chunkidx++,reqindex++){
//int debug=0;
//if(z>=38) debug=1;
int p = procmap[chunkidx];
//if(debug) printf("\tp=%u\n",p);
int *idims = gdims + chunkindex[chunkidx];
//if(debug) printf("\tdims(%u)=%u:%u:%u\n",chunkidx,idims[0],idims[1],idims[2]);
int *iorigin = gorigins + chunkindex[chunkidx];
//if(debug) printf("\torigin=%u:%u:%u\n",iorigin[0],iorigin[1],iorigin[2]);
MPI_Status stat;
//printf("waitfor(%u):MPI_Request[%u]=%u\n",z,reqindex,req[reqindex]);
// slicewait.start();
comm.wait(req[reqindex],stat); // frees request object too
//slicewait.stop();
// get information out of the status object!!!!
//slicecollect.start();
int offset = iorigin[1]*globaldims[0] + iorigin[0];
for(int idx=0,pos=offset,j=0;j<idims[1];j++,pos+=(globaldims[0]-idims[0]))
for(int i=0;i<idims[0];i++)
destbuffer[pos++]=slicebuffer[idx++];
// slicecollect.stop();
}
sliceallwait.stop();
// proclist array
}
MPIO::MPIO(IObase *io,MPIcomm &c):file(io),comm(c),slicemap(0),procmap(0),chunkindex(0),chunkindexcount(0){
for(int i=0;i<3;i++) globaldims[i]=localdims[i]=localorigin[i]=0;
int *rootarray;
myid = comm.rank();
//printf("my rank = %u\n",myid);
if(!myid) rootarray = new int[comm.numProcs()];
int rootscale = file?1:0;
//printf("rootscale[pid=%u] = %u\n",myid,rootscale);
comm.gather(0,1,&rootscale,rootarray); // gather to 0
globaldims[0]=globaldims[1]=globaldims[2]=0;
if(!myid){
for(int count=0,p=0;p<comm.numProcs();p++){
//printf("rootarray[%u]=%u\n",p,rootarray[p]);
if(rootarray[p]) count++;
}
if(count<=0 || count>1){
fprintf(stderr,"MPIO only accepts a single IO node right now.\n");
fprintf(stderr,"You chose %u io nodes\n",count);
fprintf(stderr,"I will select only the first one\n");
}
for(root=-1,p=0;p<comm.numProcs() && root<0;p++)
if(rootarray[p]) root=p;
delete rootarray;
}
comm.bcast(0,root); // now everyone knows root
//printf("broadcasting root node = %u\n",root);
}
MPIO::~MPIO(){
char buffer[32];
sprintf(buffer,"PE(%3u)",comm.rank());
if(myid==root) {
sprintf(buffer+7,"slicewait"); slicewait.print(buffer);
sprintf(buffer+7,"slicewrite"); slicewrite.print(buffer);
sprintf(buffer+7,"slicecollect"); slicecollect.print(buffer);
//sprintf(buffer+7,"sliceselect"); sliceselect.print(buffer);
sprintf(buffer+7,"sliceallwait"); sliceallwait.print(buffer);
}
}
void MPIO::setLocalDims(int rank,int *origin, int *dims){
if(rank!=3) perror("MPIO is only for 3D IO for now (sorry)!");
if(isRoot()){
gorigins = new int[3*comm.numProcs()];
gdims = new int[3*comm.numProcs()];
}
else gdims=gorigins=0; // null for everyone else
for(int i=0;i<3;i++) {
//printf("localdims-n-origin[%u] = %u:%u\n",i,dims[i],origin[i]);
localdims[i]=dims[i];
localorigin[i]=origin[i];
}
comm.gather(root,3,localdims,gdims);
comm.gather(root,3,localorigin,gorigins);
//if(isRoot()) for(i=0;i<3*comm.numProcs();i++)
//printf("gdims-n-origin[%u] = %u:%u\n",i,gdims[i],gorigins[i]);
globaldims[0]=globaldims[1]=globaldims[2]=0;
if(isRoot())
for(int p=0,last=3*comm.numProcs();p<last;p++){
//printf("globaldims[%u]: gdims[%u]+gorigins[%u] = %u\n",
// p%3,p,p,gdims[p],gorigins[p]);
if(globaldims[p%3] < gdims[p]+gorigins[p])
globaldims[p%3] = gdims[p]+gorigins[p];
}
// rebroadcast globaldims to all PE's
comm.bcast(root,3,globaldims);
// do we bcast the layout?
// Now organize the info hierarchially by Z (index into)
if(isRoot()){
if(slicemap){
delete slicemap;
delete procmap;
delete chunkindex;
delete chunkindexcount;
}
//puts("(1)");
slicemap = new int[globaldims[2]+1];
chunkindexcount = new int[globaldims[2]+1];
{
for(int thickness=0,chunkcount=0,z=0;z<globaldims[2];z+=thickness){
for(int p=0;p<comm.numProcs();p++){
// count slices
int dindex = 3*p;
int *idims = gdims + dindex;
int *iorigin = gorigins + dindex;
int iz = z-iorigin[2];
if(iz<0 || iz>=idims[2]) continue;
thickness = gdims[2]; // assumes uniform domain decomp in Z
chunkcount++;
}
}
// procmap = new int[chunkcount];
//chunkindex = new int[chunkcount]; // points into iorigin/idims
// chunkindexcount = chunkcount;
}
procmap = new int[comm.numProcs()];
chunkindex = new int [comm.numProcs()];
// printf("(2): chunkindexcount = %u\n",chunkindexcount);
for(int thickness=0,mapcount=0,chunkcount=0,z=0;z<globaldims[2];z+=thickness){
for(int p=0;p<comm.numProcs();p++){
// count slices
int dindex = 3*p;
int *idims = gdims + dindex;
int *iorigin = gorigins + dindex;
int iz = z-iorigin[2];
if(iz<0 || iz>=idims[2]) continue;
thickness = gdims[2];
procmap[chunkcount] = p;
chunkindex[chunkcount] = dindex; // point into iorogin/idims
//printf("chunkcount=%u chunkindex=%u\n",chunkcount,chunkindex[chunkcount]);
chunkcount++;
}
}
//puts("(3)");
for(chunkcount=0,z=0;z<globaldims[2];){
int thickness=0;
for(int levelcount=0,p=0;p<comm.numProcs();p++){
// count slices
int dindex = 3*p;
int *idims = gdims + dindex;
int *iorigin = gorigins + dindex;
int iz = z-iorigin[2];
if(iz<0 || iz>=idims[2]) continue;
thickness = gdims[2];
levelcount++; chunkcount++;
}
for(int pz=0;pz<thickness && z<=globaldims[2];pz++,z++){
chunkindexcount[z]=levelcount;
slicemap[z] = chunkcount - levelcount; // selects chunkindex from slicenum
//printf("slicemap[%u]= %u\n",z,slicemap[z]);
}
}
//printf("last map count = %u\n",z);
//slicemap[z] = chunkcount;
}
}
int MPIO::write(IObase::DataType type,int rank,int *dims,void *data){
int recalc_layout=0;
if(rank!=3) recalc_layout=1;
for(int i=0;i<rank;i++) if(localdims[i]!=dims[i]) recalc_layout=1;
if(recalc_layout) setLocalDims(rank,localorigin,dims);
switch(type){
case IObase::Float32:
write((float*)data);
break;
case IObase::Float64:
default:
break;
}
return 1; // for now, no error checking
}
void MPIO::write(float *data){
MPI_Request *sendreq = new MPI_Request[localdims[2]];
typedef float *floatP;
int sliceindex=0;
floatP slicebuffer[3]; // double-buffer the slices as they arrive
typedef MPI_Request *MPI_RequestP;
MPI_RequestP recvreq[2];
// the third one is a scratch buffer for reorganizing the slice
if(isRoot()){
for(int i=0;i<3;i++) slicebuffer[i] = new float[globaldims[0]*globaldims[1]];
for(i=0;i<2;i++) recvreq[i] = new MPI_Request[comm.numProcs()];
}
if(isRoot()){
// puts("request initial slice");
requestSlice(0,slicebuffer[0],recvreq[0]);
}
// not good! Everyone is sending a slice here!
sendSlice(0,data,sendreq);
// hide latency behind the costly io->reserveChunk operation
if(isRoot()) file->reserveStream(IObase::Float32,3,globaldims);
int writecount=0;
for(int i=1;i<globaldims[2];i++){
comm.barrier(); // clearly not!
if(isRoot())
requestSlice(i,slicebuffer[i%2],recvreq[i%2]);
sendSlice(i,data,sendreq); // jms
if(isRoot()) {
waitForSlice(i-1,slicebuffer[(i-1)%2],slicebuffer[2],recvreq[(i-1)%2]);
writecount+=globaldims[0]*globaldims[1];
slicewrite.start();
file->writeStream(slicebuffer[2],globaldims[0]*globaldims[1]);
slicewrite.stop();
}
}
if(isRoot()){
waitForSlice(i-1,slicebuffer[(i-1)%2],slicebuffer[2],recvreq[(i-1)%2]);
file->writeStream(slicebuffer[2],globaldims[0]*globaldims[1]);
writecount+=globaldims[0]*globaldims[1];
// do requestfree (double-buffered request free-s)
for(i=0;i<3;i++) delete slicebuffer[i]; // free everything up
for(i=0;i<2;i++) delete recvreq[i];
}
MPI_Status stat;
for(i=0;i<localdims[2];i++) comm.wait(sendreq[i],stat);
delete sendreq;
}
|
