tperf.hoc

/* File is constructed in perfrun.hoc by printing the tperf for rank 0
using p_perf where the vector tperf is retrieved after a psolve via
	tdat_.x[5] = pnm.pc.send_time(4, chist, tperf)
The first item in the file after the header is the size of tperf
See bgpdma.cpp nrn_bgp_receive_time for the meaning of the first arg. In
particular, 4 returns the number of extra conservation checks, fills chist
with the histogram of number of intervals using 0-11 conservation checks,
and fills tperf with itbuf items with the itbuf+1 item being the DCMF_Tick().
The items are in sets of DCMF_Timebase() as indicated below.
It seems itbuf is initialized only at launch and not in bgp_dma_init (which
would allow multiple runs per launch).
All the TBUF fills take place in bgp_dma_receive
Given the several multisend methods we are using and the variation in
network parameters, it is a good idea to prepend run information to the file
*/
// allgather 201 sets of
// enter nrn_spike_exchange_compressed
// another immediately
// after barrier
// after nrnmpi_spike_exchange_compressed (allgather, allgatherv)
// 0
// 0
// #sent (cells that spiked
// #received
// exit nrn_spike_exchange_compressed

// multisend 401 sets of
// enter bgp_dma_receive
// after DCMF_Messager_advance
// after barrier
// after conserve
// ncons
// #cells_that_fired
// #sent (cell spiked * fanout)
// #received
// if ENQUEUE == 1 then incoming spikes are saved in buffer, TBUF, then to queue
// exit bgp_dma_receive

{load_file("nrngui.hoc")}
objref td[1]
strdef fname
proc rd() {local i, j, k, size, toff, tm, methodinfo \
  localobj f, indices, s
	f = new File()
	f.chooser("r", "Read", "*.dat")
	if (numarg() == 1) {
		fname = $s1
	}else{
		if (!f.chooser()){
			quit()
		}
		fname = f.getname()
	}
	f.ropen(fname)
	// read header
	s = new String()
	f.gets(s.s)
	nver=0
	sscanf(s.s, "%*[^(](%d)", &nver)	
	f.gets(s.s)
	methodinfo = f.scanvar()
	method = methodinfo%4
	enqueue=int(methodinfo/32)
	if (method == 0) { enqueue = 0 }
	n_bgp_interval = (int(methodinfo/4)%2) + 1
printf("method=%d n_bgp_interval=%d enqueue=%d\n", method, n_bgp_interval,enqueue)
	while(1) {
		f.gets(s.s)
		if (strcmp(s.s,"@@@end header@@@\n") == 0) { break }
	}

	if (nver < 519) {
		nd = 8 + (enqueue == 1)
		enq1 = 7
	}else{
		nd = 9 + (enqueue == 1)
		enq1 = 8
	}
	objref td[nd]

	// read data
	size = (f.scanvar()-1)/nd
	print size
	for i=0, nd-1 {td[i] = new Vector(size)}
	for i = 0, size-1 {
		for j=0,nd-1 {
			td[j].x[i] = f.scanvar()
		}
	}
	ts = f.scanvar()
	tm = 2^32
	toff = td[0].x[0]
	indices = new Vector()
	indices.append(0, 1, 2, 3, enq1)
	if (enqueue == 1) indices.append(nd-1)
	for k=0, indices.size-1 {
		j = indices.x[k]
		// all vectors start at 0
		td[j].sub(toff)
		// take wrap around into account so vectors increase monotonically
		for i=1, size-1 {
			while (td[j].x[i] < td[j].x[i-1]) {
				td[j].x[i] += tm
			}
		}
		//td[j].mul(ts)
	}
	print ts
	f.close	
}
//rd(1)

obfunc combine() { local i localobj x
	x = new Vector()
printf("combine size=%d\n", $o1.size)
	for (i=0; i < $o1.size; i += 2) {
		x.append($o1.x[i] + $o1.x[i+1])
	}
	return x
}

objref g, g1, g2, yy[7], name[7]
proc gr() {local i, a, b, n localobj tt, xx, s1
	g1 = new Graph(0)
	g1.view(0,0,200,12, 100, 100, 200, 500)
	//g2 = new Graph()
	g1.size(0,200,0,15)
	//g2.size(0,40,0,6e6)
	g1.label(.3, .9, fname, 2)
	//g2.label(.5, .9, fname, 2)
	g = g1
	// tt is the ticks between beginning of next spike exchange and
	// end of previous spike exchange. Since the first spike exchange
	// takes place during finitialize, the ith tt interval is the
	// interval computation time between the ith and i+1st spike exchange.
	// Note that the first interval contains some interpreter and other
	// overhead since it is the time between the return of the finitialize
	// call to nrn_spike_exchange, and the entry to nrn_spike_exchange
	// from pc.psolve
	tt = new Vector()
	for i=0, td[nd-1].size-2 tt.append(td[0].x[i+1] - td[nd-1].x[i])
	n = tt.size
	if (n > 300) {
		g1.size(0,400,0,6)
	}else{
		g1.size(0,200,0,12)
	}
	xx = new Vector(n)
	// compute
	yy[5] = tt.div(1e6)
	name[5] = new String("interval compute")

	xx = yy[5].c
	a=xx.sum*ts*1e6
	//labl(xx)
	s1 = xx.c // drawn at end below
	
{//  if (n > 300) {
	// first messager advance
	yy[0] = td[1].c.sub(td[0]).div(1e6).c(1,n)
	name[0] = new String("Messager advance")
	xx = xx.add(yy[0])
	//labl(xx)
	//xx.line(g, 3, 1)
  }

	// barrier
	yy[1] = td[2].c.sub(td[1]).div(1e6).c(1,n)
	name[1] = new String("barrier")
	// conserve (or allgather+allgatherv)
	yy[2] = td[3].c.sub(td[2]).div(1e6).c(1,n)
	name[2] = new String("conserve")
	// incoming spikes saved to buffer (or allgather receive enqueue)
	yy[3] = td[enq1].c.sub(td[3]).div(1e6).c(1,n)
	name[3] = new String("spike buffer")
	if (enqueue == 1) {
		// buffered incoming spikes enqueued
		yy[4] = td[enq1+1].c.sub(td[enq1]).div(1e6).c(1,n)
		name[4] = new String("enqueuing")
	}else{
		yy[4] = yy[3].c.fill(0)
		name[4] = new String("nothing")
	}
	// after barrier to after barrier
	yy[6] = new Vector()
	for i=0, td[2].size-2 yy[6].append(td[2].x[i+1] - td[2].x[i])
	yy[6].div(1e6)
	name[6] = new String("barrier to barrier")

	xx = yy[0].c.fill(0)
	for case (&i, 5, 0, 1, 2, 3, 4) {
	//for case (&i, 2, 3, 4, 5, 0, 1) {
		xx.add(yy[i])
		labl(xx, i)
		xx.line(g, i+1, 1)
	}
/*		
	xx = xx.add(yy[1])
	labl(xx)
	xx.line(g, 2, 1)

	xx = xx.add(yy[2])
	labl(xx)
	xx.line(g, 5, 1)

	xx = xx.add(yy[3])
	labl(xx)
	xx.line(g, 6, 1)

	if (enqueue == 1) {
		xx = xx.add(yy[4])
		labl(xx)
		print xx.line(g, 7, 1).sum*ts*1e6
	}
*/
	labl(yy[6], 6)
	print yy[6].line(g, 8, 1).sum*ts*1e6

	// tt (the interval computation time)
//	s1.line(g, 1, 1).sum*ts*1e6

//	print "total = ", a+b
//	g.exec_menu("View = plot")
}

proc labl() {local x  localobj s
	s = new String()
	x = $o1.sum*ts*1e6
	printf("%3.2fs %s\n", x, name[$2].s)
	sprint(s.s, "%3.2fs %s", x, name[$2].s)
	g1.label($o1.size*1.02, $o1.c($o1.size-101, $o1.size-1).mean,s.s,1,1,0,.5,1)
}
//gr()

objref list
list = new List()
proc again() {
	list.append(g1) list.append(g2)
	rd()
	gr()
}

proc ag() {
	list.append(g1) list.append(g2)
	rd($s1)
	gr()
}

/*
ag("nrn-results/1kconn/results_256_8/t0perf18.1k.8K.dat")
ag("nrn-results/1kconn/results_256_8/t2perf18.1k.8K.dat")
ag("nrn-results/1kconn/results_256_8/t6perf18.1k.8K.dat")

ag("nrn-results/1kconn/results_256_16/t0perf18.0.16384.dat")
ag("nrn-results/1kconn/results_256_16/t2perf18.0.16384.dat")
ag("nrn-results/1kconn/results_256_16/t6perf18.0.16384.dat")

ag("nrn-results/1kconn/results_256_32/t0perf18.0.32768.dat")
ag("nrn-results/1kconn/results_256_32/t2perf18.0.32768.dat")
ag("nrn-results/1kconn/results_256_32/t6perf18.0.32768.dat")
*/