merge

ericvh committed 16 years ago

Commit f68939a2 parents 6083349 da88d0b5 files · +67 −34Browse files

Unified · Split

5 files

sys/src/cmd/rompi/mkfile +1 −0
sys/src/cmd/rompi/reduce.c +17 −5
sys/src/cmd/rompi/rompi.c +18 −17
sys/src/cmd/rompi/stupidreduce.c +29 −10
sys/src/cmd/rompi/torus.c +2 −2

@@ -13,6 +13,7 @@ DIRS=\ · 12 unmodified lines
13	13	OTHEROFILES=\
14	14	torus.$O\
15	15	rompi.$O\
	16	+ support.$O\
16	17
17	18	HFILES=mpi.h
18	19

sys/src/cmd/rompi/reduce.c+17−5

@@ -5,6 +5,7 @@ · 4 unmodified lines
5	5	#include "mpi.h"
6	6
7	7	/* tested may 17 2010 and works */
	8	+u64int tbget(void);
8	9	extern int torusdebug, rompidebug;
9	10	extern int x, y, z, xsize, ysize, zsize;
10	11	int xyztorank(int x, int y, int z);
@@ -16,20 +17,31 @@ main(int argc, char **argv) · 5 unmodified lines
16	17	{
17	18	int node, nproc;
18	19	unsigned int sum[1], nsum[1];
	20	+ int i;
	21	+ vlong startnsec, stopnsec;
	22	+ int iter = 100;
19	23
20	24	MPI_Init(&argc,&argv);
21	25	MPI_Comm_rank(MPI_COMM_WORLD, &node);
22	26	MPI_Comm_size(MPI_COMM_WORLD, &nproc);
23	27
24		− print("%lld Hello World from Node %d of %d\n", nsec(),node, nproc);
25		−
26		− sum[0] = node;
	28	+ print("%lld Hello World from Node %d of %d\n", tbget(),node, nproc);
27	29
28	30	if (0 && node)
29	31	rompidebug \|= 4;
30		− MPI_Reduce (sum, nsum, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
	32	+ for (i = 0; i < iter; i++){
	33	+ if (i == 1)
	34	+ startnsec = tbget();
	35	+ sum[0] = node;
	36	+ MPI_Reduce (sum, nsum, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
	37	+// if (! node)
	38	+// print("%d\n", i);
	39	+ }
	40	+ stopnsec = tbget();
31	41
32		− print("%lld Node %d (%d, %d, %d) computes sum as %d and MPI says sum is %d\n", nsec(), node, x, y, z, ((nproc-1)*nproc)/2, nsum[0]);
	42	+ print("%lld Node %d (%d, %d, %d) computes sum as %d and MPI says sum is %d\n", tbget(), node, x, y, z, ((nproc-1)*nproc)/2, nsum[0]);
	43	+ if (! node)
	44	+ print("%lld nsec for %d iterations\n", stopnsec-startnsec, iter);
33	45	print("%lld %d: Finalize\n", nsec(), node);
34	46	MPI_Finalize();
35	47	}

sys/src/cmd/rompi/rompi.c+18−17

@@ -40,6 +40,7 @@ void torusinit(int *pmyproc, const int pnprocs); · 39 unmodified lines
40	40	void ranktoxyz(int rank, int x, int y, int *z);
41	41	int xyztorank(int x, int y, int z);
42	42	extern int x, y, z, xsize, ysize, zsize;
	43	+u64int tbget(void);
43	44
44	45	void
45	46	buf_add(struct bufpacket *new)
@@ -282,7 +283,7 @@ int reduce_end ( void buf, int num, MPI_Datatype datatype, int /root*/, · 236 unmodified lines
282	283	torustag[TAGtag] = reducetag;
283	284	torustag[TAGsource] = myproc;
284	285
285		− if(rompidebug&4) print("%lld reduce_end: %d(%d, %d, %d)\n", nsec(), myproc, x, y, z);
	286	+ if(rompidebug&4) print("%lld reduce_end: %d(%d, %d, %d)\n", tbget(), myproc, x, y, z);
286	287	if (myproc) {
287	288	/* who we get it from depends on who we are. */
288	289	/* this shows some bad design. We compute x,y,z, turn it into rank,
@@ -298,9 +299,9 @@ int reduce_end ( void buf, int num, MPI_Datatype datatype, int /root*/, · 9 unmodified lines
298	299	else if (x)
299	300	fromz = fromy = fromx = 0;
300	301	fromrank = xyztorank(fromx, fromy, fromz);
301		− if(rompidebug&4) print("%lld reduce_end: %d(%d,%d,%d): wait from (%d, %d, %d)\n", nsec(), myproc, x, y, z, fromx, fromy, fromz);
	302	+ if(rompidebug&4) print("%lld reduce_end: %d(%d,%d,%d): wait from (%d, %d, %d)\n", tbget(), myproc, x, y, z, fromx, fromy, fromz);
302	303	MPI_Recv(buf, 1, MPI_INT, fromrank, reducetag, MPI_COMM_WORLD, status);
303		− if(rompidebug&4) print("%lld reduce_end: %d: Got from %d\n", nsec(), myproc, fromrank);
	304	+ if(rompidebug&4) print("%lld reduce_end: %d: Got from %d\n", tbget(), myproc, fromrank);
304	305	}
305	306
306	307	/* OK, we got it, we just need to do one send to our "axis" (of evil?) */
@@ -311,24 +312,24 @@ int reduce_end ( void buf, int num, MPI_Datatype datatype, int /root*/, · 4 unmodified lines
311	312	*/
312	313	if (z) {
313	314	} else if (y) {
314		− if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", nsec(), myproc, x, y, z, x, y, zsize-1);
	315	+ if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", tbget(), myproc, x, y, z, x, y, zsize-1);
315	316	torustag[TAGsource] = myproc;
316	317	torussend(buf, count, x, y, zsize-1, 1, torustag, sizeof(torustag));
317	318	}
318	319	else if (x) {
319	320	/* send down BOTH our y and z axis */
320	321	torustag[TAGsource] = myproc;
321		− if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", nsec(), myproc, x, y, z, x, ysize-1, 0);
	322	+ if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", tbget(), myproc, x, y, z, x, ysize-1, 0);
322	323	torussend(buf, count, x, ysize-1, 0, 1, torustag, sizeof(torustag));
323		− if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", nsec(), myproc, x, y, z, x, y, zsize-1);
	324	+ if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", tbget(), myproc, x, y, z, x, y, zsize-1);
324	325	torussend(buf, count, x, y, zsize-1, 1, torustag, sizeof(torustag));
325	326	} else {
326	327	/* send down x, y, z */
327		− if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", nsec(), myproc, x, y, z, xsize-1, 0, 0);
	328	+ if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", tbget(), myproc, x, y, z, xsize-1, 0, 0);
328	329	torussend(buf, count, xsize-1, 0, 0, 1, torustag, sizeof(torustag));
329		− if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", nsec(), myproc, x, y, z, x, ysize-1, 0);
	330	+ if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", tbget(), myproc, x, y, z, x, ysize-1, 0);
330	331	torussend(buf, count, x, ysize-1, 0, 1, torustag, sizeof(torustag));
331		− if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", nsec(), myproc, x, y, z, x, y, zsize-1);
	332	+ if(rompidebug&4) print("%lld reduce_end %d(%d, %d, %d): send to (%d, %d, %d)\n", tbget(), myproc, x, y, z, x, y, zsize-1);
332	333	torussend(buf, count, x, y, zsize-1, 1, torustag, sizeof(torustag));
333	334	}
334	335
@@ -344,7 +345,7 @@ opsum(void dst, void src, int) · 9 unmodified lines
344	345	int *is = src;
345	346	int *id = dst;
346	347	int tmp = *is;
347		− if (rompidebug & 4) print("%lld %d(%d,%d,%d): SUM %d+%d=%d\n", nsec(), myproc, x,y,z,id, is, *id+tmp);
	348	+ if (rompidebug & 4) print("%lld %d(%d,%d,%d): SUM %d+%d=%d\n", tbget(), myproc, x,y,z,id, is, *id+tmp);
348	349	*id += tmp;
349	350	return *id;
350	351	}
@@ -397,7 +398,7 @@ int MPI_Reduce ( void sendbuf, void recvbuf, int count, · 46 unmodified lines
397	398	if (z) {
398	399	torank = xyztorank(x,y,0);
399	400	if (rompidebug&4)
400		− print("%lld %d:(%d,%d,%d) sends %d to %d\n", nsec(), myproc, x, y, z, myproc, torank);
	401	+ print("%lld %d:(%d,%d,%d) sends %d to %d\n", tbget(), myproc, x, y, z, myproc, torank);
401	402	MPI_Send(tmp, 1, datatype, torank, 1, comm);
402	403	} else {
403	404	/* gather up all our z >0 data */
@@ -407,14 +408,14 @@ int MPI_Reduce ( void sendbuf, void recvbuf, int count, · 3 unmodified lines
407	408	op(tmp, recvbuf, count);
408	409	}
409	410	if (rompidebug&4)
410		− print("%lld AFTER Z: %d(%d, %d, %d): %d\n", nsec(), myproc, x, y, z, sum[0]);
	411	+ print("%lld AFTER Z: %d(%d, %d, %d): %d\n", tbget(), myproc, x, y, z, sum[0]);
411	412
412	413	/* y? y o y? */
413	414	if (y) {
414	415	torank = xyztorank(x,0,0);
415	416	/* send to our parent */
416	417	if (rompidebug&4)
417		− print("%lld %d:(%d,%d,%d) sends %d to %d\n", nsec(), myproc, x, y, z, sum[0], torank);
	418	+ print("%lld %d:(%d,%d,%d) sends %d to %d\n", tbget(), myproc, x, y, z, sum[0], torank);
418	419	MPI_Send(tmp, 1, datatype, torank, 1, comm);
419	420	} else {
420	421	/* This is the vector along the X axis. All myprocs, including (0,0,0), gather up along
@@ -429,11 +430,11 @@ int MPI_Reduce ( void sendbuf, void recvbuf, int count, · 8 unmodified lines
429	430	op(tmp, recvbuf, count);
430	431	}
431	432	if (rompidebug&4)
432		− print("%lld %d:(%d,%d,%d) gets %d accum %d\n", nsec(), myproc, x, y, z, nsum[0], sum[0]);
	433	+ print("%lld %d:(%d,%d,%d) gets %d accum %d\n", tbget(), myproc, x, y, z, nsum[0], sum[0]);
433	434	if (x) {
434	435	/* send to our parent */
435	436	if (rompidebug&4)
436		− print("%lld %d:(%d,%d,%d) sends %d to %d\n", nsec(), myproc, x, y, z, sum[0], torank);
	437	+ print("%lld %d:(%d,%d,%d) sends %d to %d\n", tbget(), myproc, x, y, z, sum[0], torank);
437	438	MPI_Send(tmp, 1, datatype, torank, 1, comm);
438	439	} else {
439	440	/* sum contains the sum from our y axis above */
@@ -441,7 +442,7 @@ int MPI_Reduce ( void sendbuf, void recvbuf, int count, · 1 unmodified line
441	442	MPI_Recv(recvbuf, 1, datatype, i, 1, comm, &status);
442	443	op(tmp, recvbuf, count);
443	444	if (rompidebug&4)
444		− print("%lld %d:(%d,%d,%d) recv %d to %d\n", nsec(), myproc, x, y, z, sum[0], i);
	445	+ print("%lld %d:(%d,%d,%d) recv %d to %d\n", tbget(), myproc, x, y, z, sum[0], i);
445	446	}
446	447	}
447	448	}
@@ -450,7 +451,7 @@ int MPI_Reduce ( void sendbuf, void recvbuf, int count, · 2 unmodified lines
450	451	if (! myproc){
451	452	memmove(recvbuf, tmp, nbytes);
452	453	if (rompidebug&4)
453		− print("%lld reduce: %d(%d,%d,%d): gather done, send SUM %d\n", nsec(), myproc, x, y, z, sum[0]);
	454	+ print("%lld reduce: %d(%d,%d,%d): gather done, send SUM %d\n", tbget(), myproc, x, y, z, sum[0]);
454	455	}
455	456
456	457	return reduce_end (recvbuf, count, datatype, root, comm, &status);

sys/src/cmd/rompi/stupidreduce.c+29−10

@@ -6,21 +6,14 @@ · 5 unmodified lines
6	6
7	7	/* tested may 17 2010 and works */
8	8	extern int torusdebug;
	9	+u64int tbget(void);
	10	+int node, nproc;
9	11
10	12	void
11		−main(int argc, char **argv)
	13	+stupidreduce(unsigned int *sum)
12	14	{
13		− int node, nproc;
14		− unsigned int sum[1];
15	15	MPI_Status status;
16	16	int i;
17		−
18		− MPI_Init(&argc,&argv);
19		− MPI_Comm_rank(MPI_COMM_WORLD, &node);
20		− MPI_Comm_size(MPI_COMM_WORLD, &nproc);
21		−
22		− print("Hello World from Node %d of %d\n",node, nproc);
23		−
24	17	if (node) {
25	18	sum[0] = node;
26	19	MPI_Send(sum, 1, MPI_INT, 0, 1, MPI_COMM_WORLD);
@@ -36,7 +29,33 @@ main(int argc, char **argv) · 9 unmodified lines
36	29	MPI_Send(sum, 1, MPI_INT, i, 1, MPI_COMM_WORLD);
37	30	}
38	31	}
	32	+}
	33	+
	34	+void
	35	+main(int argc, char **argv)
	36	+{
	37	+ unsigned int sum[1];
	38	+ int i;
	39	+ vlong startnsec, stopnsec;
	40	+ int iter = 100;
	41	+
	42	+ MPI_Init(&argc,&argv);
	43	+ MPI_Comm_rank(MPI_COMM_WORLD, &node);
	44	+ MPI_Comm_size(MPI_COMM_WORLD, &nproc);
	45	+
	46	+ print("Hello World from Node %d of %d\n",node, nproc);
	47	+
	48	+ for(i = 0; i < iter; i++){
	49	+ if (i == 1)
	50	+ startnsec = tbget();
	51	+ sum[0] = node;
	52	+ stupidreduce(sum);
	53	+ }
	54	+ stopnsec = tbget();
	55	+
39	56	print("%d: Finalize\n", node);
	57	+ if (! node)
	58	+ print("%lld nsec for %d iterations\n", stopnsec-startnsec, iter-1);
40	59	MPI_Finalize();
41	60	print("Node %d computes sum as %d and MPI says sum is %d\n", node, ((nproc-1)*nproc)/2, sum[0]);
42	61	}

sys/src/cmd/rompi/torus.c+2−2

@@ -205,11 +205,11 @@ void torusinit(int *pmyproc, const int numprocs) · 204 unmodified lines
205	205	z = d[Z];
206	206
207	207	*pmyproc = (z << (lx + ly)) + (y << lx) + x;
208		− /*
	208	+
209	209	print( "%d/%d %d/%d %d/%d\n", x, xsize, y, ysize, z, zsize);
210	210	print( "numprocs %d \n", numprocs);
211	211	print("%d/%d %d/%d %d/%d\n", lx, xbits, ly, ybits, lz, zbits);
212		− */
	212	+
213	213	/* make some tables. Mapping is done as it is to maximally distributed broadcast traffic */
214	214	xyz = calloc(numprocs, sizeof(*xyz));
215	215	if (!xyz)