/**
* This static method creates the main function as described in the class comment. It first checks
* that the filter has no inputs/outputs and does not declare any I/O rates Then it creates the
* function and attaches it to the filter.
*
* @param top The single flatnode that contains the single filter of the app.
* @param executionCounts The schedule for the stream graph as calculated by the scheduler.
*/
public static void doit(FlatNode top, HashMap[] executionCounts) {
assert top.contents instanceof SIRFilter : "Error top of graph is not filter";
assert top.inputs == 0 && top.ways == 0 : "Error: Fused filter contains neighbors";
SIRFilter filter = (SIRFilter) top.contents;
assert filter.getPushInt() == 0 && filter.getPopInt() == 0 && filter.getPeekInt() == 0
: "Error: fused filter declares non-zero I/O rate(s)";
// make sure there are no push, pops or peeks...
assert CheckForCommunication.check(filter) == false
: "Error: Communication expression found in filter";
// assert !(filter instanceof SIRTwoStageFilter) :
// "Error: Fused filter is a two stage";
ExecutionCode exeCode = new ExecutionCode();
// check to see if the schedule is valid and
// set the number of times the filter fires in the init
// stage (weird)
exeCode.checkSchedule(executionCounts, filter);
// create the main function of the C code that will call
// the filter
JBlock block = exeCode.mainFunction(filter);
// create the method and add it to the filter
JMethodDeclaration mainFunct =
new JMethodDeclaration(
null,
at.dms.kjc.Constants.ACC_PUBLIC,
CStdType.Void,
Names.main,
JFormalParameter.EMPTY,
CClassType.EMPTY,
block,
null,
null);
// make the main the new work function
filter.setWork(mainFunct);
}
/**
* generate the code for the steady state loop. Inline the work function inside of an infinite
* while loop.
*
* @param filter The single fused filter of the application.
* @return A JStatement that is a while loop with the work function inlined.
*/
JStatement generateSteadyStateLoop(SIRFilter filter) {
JBlock block = new JBlock(null, new JStatement[0], null);
JBlock workBlock = (JBlock) ObjectDeepCloner.deepCopy(filter.getWork().getBody());
// add the cloned work function to the block
block.addStatement(workBlock);
// return the infinite loop
return new JWhileStatement(null, new JBooleanLiteral(null, true), block, null);
}
private boolean does_peek(int index) {
CConfig child = childConfig(index);
if (child instanceof CConfigFilter) {
CConfigFilter fc = (CConfigFilter) child;
SIRFilter filter = (SIRFilter) fc.getStream();
int f_pop = filter.getPopInt();
int f_peek = filter.getPeekInt();
if (f_peek > f_pop) return true;
}
if (child instanceof CConfigSplitJoin) {
CConfigSplitJoin sj = (CConfigSplitJoin) child;
if (sj.getPeek()) {
return true;
}
}
return false;
}
/**
* Construct the main function.
*
* @param filter The single SIRFilter of the application.
* @return A JBlock with the statements of the main function.
*/
private JBlock mainFunction(SIRFilter filter) {
JBlock statements = new JBlock(null, new JStatement[0], null);
// create the params list, for some reason
// calling toArray() on the list breaks a later pass
List paramList = filter.getParams();
JExpression[] paramArray;
if (paramList == null || paramList.size() == 0) paramArray = new JExpression[0];
else paramArray = (JExpression[]) paramList.toArray(new JExpression[0]);
// add the call to the init function
statements.addStatement(
new JExpressionStatement(
null,
new JMethodCallExpression(
null, new JThisExpression(null), filter.getInit().getName(), paramArray),
null));
// add the call to the pre(init)work function if this filter
// is a two stage..
if (filter instanceof SIRTwoStageFilter) {
SIRTwoStageFilter two = (SIRTwoStageFilter) filter;
statements.addStatement(
new JExpressionStatement(
null,
new JMethodCallExpression(
null, new JThisExpression(null), two.getInitWork().getName(), new JExpression[0]),
null));
}
// add the call to the work function
statements.addStatement(generateSteadyStateLoop(filter));
return statements;
}
public FusionInfo getFusionInfo(int from, int to) {
// check if we have precomputed
if (fusion_info[from][to] != null) return fusion_info[from][to];
long work = 0;
long work_no_penalty = 0;
int code = 0;
int data = 0;
int mult[] = new int[cont.size()];
mult[from] = 1; // init mult. of first item to 1
for (int i = from; i < to; i++) {
int push = childConfig(i).getFusionInfo().getPushInt() * mult[i];
int pop = childConfig(i + 1).getFusionInfo().getPopInt();
int common = gcd(push, pop);
int xtra = pop / common;
if (xtra > 1) {
for (int a = from; a <= i; a++) {
mult[a] = mult[a] * xtra;
}
}
mult[i + 1] = (push * xtra) / pop;
}
for (int i = from; i <= to; i++) {
CConfig child = childConfig(i);
FusionInfo fi = child.getFusionInfo();
work += fi.getWorkEstimate();
work_no_penalty += fi.getWorkEstimateNoPenalty();
code += fi.getCodeSize() * mult[i];
data += fi.getDataSize(); // * mult[i]; // removed Nov-15-2004
// removed Nov-15-2004
// if (i < to) data += fi.getOutputSize() * fi.getPushInt() * mult[i] * 2;
// add impossible unroll penalty
if (KjcOptions.unroll < mult[i]) work += fi.getWorkEstimate() / 2;
// If peek ratio is 1024 add fusion peek overhead!
if (KjcOptions.peekratio >= 1024) {
if (i > from && child instanceof CConfigFilter) {
CConfigFilter fc = (CConfigFilter) child;
SIRFilter filter = (SIRFilter) fc.getStream();
int f_pop = filter.getPopInt();
int f_peek = filter.getPeekInt();
int f_extra = f_peek - f_pop;
if (f_extra > 0) { // peek > pop
// only penalize if peek rate < 25% of (peek-pop)
// System.out.print("[ extra:"+f_extra+"("+(f_pop * mult[i] * 4)+") pop:"+f_pop+"
// m:"+mult[i]+" ]");
if (f_extra > f_pop * mult[i] * 4) {
// work += fi.getWorkEstimate()/2;
// work += 1500;
// System.out.print("penalty!!");
}
// System.out.println();
}
}
if (i > from && child instanceof CConfigSplitJoin) {
CConfigSplitJoin sj = (CConfigSplitJoin) child;
if (sj.getPeek()) {
// work += fi.getWorkEstimate()/2;
// work += 1500;
}
}
}
}
int to_push = childConfig(to).getFusionInfo().getPushInt();
int from_pop = childConfig(from).getFusionInfo().getPopInt();
int from_peek = childConfig(from).getFusionInfo().getPeekInt();
fusion_info[from][to] =
new FusionInfo(
work,
work_no_penalty,
code,
data,
from_pop * mult[from],
from_pop * mult[from] + (from_peek - from_pop),
to_push * mult[to],
childConfig(from).getFusionInfo().getInputSize(),
childConfig(to).getFusionInfo().getOutputSize());
return fusion_info[from][to];
}
private int refactor(int from, int to) {
// System.out.println("refactor from:"+from+" to:"+to);
int new_cuts = 0;
// find multiplicity of filters
int mult[] = new int[cont.size()];
mult[from] = 1; // init mult. of first item to 1
for (int i = from; i < to; i++) {
int push = childConfig(i).getFusionInfo().getPushInt() * mult[i];
int pop = childConfig(i + 1).getFusionInfo().getPopInt();
int common = gcd(push, pop);
int xtra = pop / common;
if (xtra > 1) {
for (int a = from; a <= i; a++) {
mult[a] = mult[a] * xtra;
}
}
mult[i + 1] = (push * xtra) / pop;
}
for (int i = from + 1; i <= to; i++) {
CConfig child = childConfig(i);
// If peek ratio is 1024 -- try to refactor partition
if (KjcOptions.peekratio >= 1024 && child instanceof CConfigFilter) {
CConfigFilter fc = (CConfigFilter) child;
SIRFilter filter = (SIRFilter) fc.getStream();
int f_pop = filter.getPopInt();
int f_peek = filter.getPeekInt();
int f_extra = f_peek - f_pop;
if (f_extra > 0) { // peek > pop
// double amortize_break = f_extra / ClusterBackend.getExecCounts(filter);
// double amortize_fuse = f_extra / mult[i];
// System.out.print("[refactor extra:"+f_extra+"("+(f_pop * mult[i] * 4)+") pop:"+f_pop+"
// m:"+mult[i]+" ]");
// break if peek rate < 25% of (peek-pop)
// if (f_extra > f_pop * mult[i] * 4) {
// System.out.print("Cutting partition!!");
greedyCuts[i - 1] = true;
new_cuts++;
// }
/*
else if (amortize_break * 2 < amortize_fuse) {
// break if amortize_break * 2 < amortize_fuse
//System.out.print("Cutting partition!!");
greedyCuts[i-1] = true;
new_cuts++;
}
*/
// System.out.println();
}
}
if (i > from && child instanceof CConfigSplitJoin) {
CConfigSplitJoin sj = (CConfigSplitJoin) child;
if (sj.getPeek()) {
// System.out.print("Cutting partition!!");
greedyCuts[i - 1] = true;
new_cuts++;
// work += fi.getWorkEstimate()/2;
// work += 1500;
}
}
}
return new_cuts;
}