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
|
#include <stdio.h>
#include "MPIO.hh"
MPIO::MPIO(IEEEIO *io, MPIcomm &c):file(io),comm(c),sendreq(0){
for(int i=0;i<3;i++) globaldims[i]=localdims[i]=localorigin[i]=0;
myid=comm.rank();
int *rootarray=0;
if(!myid) rootarray=new int[comm.numProcs()];
int hasfile= file?1:0;
comm.gather(0,1,&hasfile,rootarray);
if(!myid){
int count,p;
for(count=0,p=0;p<comm.numProcs();p++){
if(rootarray[p]) { count++; if(count<=1) root=p; }
}
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(int i=0;i<3;i++) slicebuffer[i]=0;
for(int i=0;i<2;i++) recvreq[i]=0;
comm.bcast(0,root); // tell everyone which proc is root
if(rootarray) {delete rootarray; rootarray=0;}
}
MPIO::~MPIO(){}
void MPIO::setLocalDims(int rank,int *origin, int *dims){
if(rank!=3) perror("MPIO is only for 3D IO for now (sorry)!");
if(sendreq) { puts("Bad move!!! Die!!!"); exit(0); }
// now check to see if anything has changed
// actually we still need to do this collection because we
// don't know if something might have changed on some other processor
// URK...
allorigins = new int[3*comm.numProcs()];
alldims = new int[3*comm.numProcs()];
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];
}
for(int i=0;i<comm.numProcs()*3;i++) alldims[i]=allorigins[i]=-1;
comm.gather(0,3,localdims,alldims);
comm.gather(0,3,localorigin,allorigins);
comm.bcast(0,comm.numProcs()*3,alldims);
comm.bcast(0,comm.numProcs()*3,allorigins);
// now everyone figures out the global dimensions
// from max offset+dims
for(int p=0,last=3*comm.numProcs();p<last;p++){
// if(myid) fprintf(stderr,"Proc[%u] p=%d p%3=%d size=%d\n\to:d = %d:%d\n",
// if(myid) fprintf(stderr,"Proc[%u] p=%d o:d = %d:%d\n",
// myid,p,/* p%3,alldims[p]+allorigins[p],*/
// allorigins[p],alldims[p]);
if(globaldims[p%3] < alldims[p]+allorigins[p])
globaldims[p%3] = alldims[p]+allorigins[p];
}
// nothing to rebroadcast here...
//fprintf(stderr,"PE(%u): setLocalDims: globaldims=%u,%u,%d\n",
// myid,globaldims[0],globaldims[1],globaldims[2]);
// lets re-compute the requests and stuff
if(sendreq) delete sendreq; // clean out the old stuff
sendreq = new MPI_Request[localdims[2]];
if(isRoot()){
for(int i=0;i<3;i++) {
if(slicebuffer[i]) delete slicebuffer[i];
slicebuffer[i] = new float[globaldims[0]*globaldims[1]*3+27]; // padded (extremely)
}
// printf("Slicebuffer size=%u\n",globaldims[0]*globaldims[1]*2);
for(int i=0;i<2;i++) {
if(recvreq[i]) delete recvreq[i];
recvreq[i] = new MPI_Request[comm.numProcs()*4+7]; // padded
}
}
}
void MPIO::asyncWrite(float *data){
//puts("Begin AsyncWrite");
// get things started by getting root to request initial slice
if(isRoot()){
requestSlice(0,slicebuffer[0],recvreq[0]);
}
// comm.print("Requested First Slice");
// now everyone sends their slice over
sendSlice(0,data,sendreq); // separate request for each slice
//comm.print("Sent First slice");
if(isRoot()){
file->reserveStream(IObase::Float32,3,globaldims);
}
for(int i=1;i<globaldims[2];i++){
if(isRoot()) // if I'm root, then make request for nex slice
requestSlice(i,slicebuffer[i%2],recvreq[i%2]);
sendSlice(i,data,sendreq); // separate sendreq for each slice
if(isRoot()) {
// waitfor slice and assemble into the slicebuffer
waitForSlice(i-1,slicebuffer[(i-1)%2],slicebuffer[2],recvreq[(i-1)%2]);
// now write assembled slice to disk stream
// can make disk buffer huge to simulate recieving multiple
// slices
//fprintf(stderr,"WriteStream Slice %u\n",i-1);
file->writeStream(slicebuffer[2],globaldims[0]*globaldims[1]);
}
}
if(isRoot()){
int n = globaldims[2]-1;
waitForSlice(n,slicebuffer[(n)%2],slicebuffer[2],recvreq[(n)%2]);
file->writeStream(slicebuffer[2],globaldims[0]*globaldims[1]);
// do requestfree (double-buffered request free-s)
//for(int i=0;i<3;i++) delete slicebuffer[i]; // free everything up
//for(int i=0;i<2;i++) delete recvreq[i];
}
//**** At end (matching deletes) ****
// fprintf(stderr,"Cleanup proc %u\n",myid);
// comm.print("Cleanup");
for(int n=0;n<localdims[2];n++){
MPI_Status stat;
comm.wait(sendreq[n],stat);
}
// perhaps should be a persistent variable
// delete sendreq;
//sendreq=0; // null it out so we know we are done
}
void MPIO::sendSlice(int z,float *data,MPI_Request *req){
int lz = z - localorigin[2]; // find local z-offset here
// filter out bogus requests
// send even redundant information though...
if(lz<0 || lz >=localdims[2]) return;
// fprintf(stderr,"\tSendSlice: Proc %u Slice %u sndindex=%u\n",
// myid,z,lz);
comm.iSend(root,z,
localdims[0]*localdims[1], // size of the data
data+localdims[0]*localdims[1]*lz,req[lz]);
}
int MPIO::requestSlice(int z,float *slicebuf, MPI_Request *req){
if(!isRoot()) return 0; // only root performs this operation
// for every processor that intersects this slice boundary,
// make a request for the contents of that slice!
int reqindex =0;
for(int offset=0,p=0;p<comm.numProcs();p++){
int idx = 3*p;
int *dims = alldims+idx;
int *origin = allorigins+idx; // need to pre-fix the origins here?
// thus to prevent mishandling of ghost zones here.
// or else have provision to support multi-reception of the same
// tile..... well actually we can do that already...
// we just have to have buckets for 2x more requests.
if(z>=origin[2] && z<origin[2]+dims[2]){
// fprintf(stderr,"\tRequestSlice: Proc %u Slice %u reqindex=%u\n",
// p,z,reqindex);
comm.iRecv(p,z,dims[0]*dims[1],slicebuf+offset,req[reqindex++]);
// fprintf(stderr,"\tSlicebuf offset=%u of size %u ++ end=%u vs sbsize=%u %u:%u @ %u:%u\n",
// offset,dims[0]*dims[1],
// offset+dims[0]*dims[1],globaldims[0]*globaldims[1]*2,
// globaldims[0],globaldims[1],dims[0],dims[1]);
offset+=dims[0]*dims[1];
}
}
return reqindex;
}
/*
Problems: Slice size is too small for slice reception with ghost zones
Also doesn't get proper offset to each slice in the slice buffer (maybe ok)
*/
void MPIO::waitForSlice(int z,float *slicebuf,
float *destbuffer,MPI_Request *req){
if(!isRoot()) return;
for(int p=0,reqindex=0;p<comm.numProcs();p++) {
MPI_Status stat;
int idx = 3*p;
int *dims = alldims+idx;
int *origin = allorigins+idx; // need to pre-fix the origins here?
// thus to prevent mishandling of ghost zones here.
if(z>=origin[2] && z<origin[2]+dims[2]){
// wiat for the data transfer to complete
//fprintf(stderr,"\tWaitSlice: Proc %u Slice %u reqindex=%u\n",
// p,z,reqindex);
comm.wait(req[reqindex++],stat); // frees request object too
// now we integrate the slices together. (data unscrambling phase)
//fprintf(stderr,"\t****Now Unscramble\n");
int offset = origin[1]*globaldims[0] + origin[0]; // initial offset for dst
for(int idx=0,pos=offset,j=0;j<dims[1];j++,pos+=(globaldims[0]-dims[0])){
for(int i=0;i<dims[0];i++){
destbuffer[pos++]=slicebuf[idx++]; // FUCK!!!! idx is questionable
}
}
// fprintf(stderr,"\t*****\n");
}
}
}
|
