const CompSlice* ParameterizeBlocking::
SetBlocking(CompSliceLocalityRegistry *anal,
const CompSliceDepGraphNode::FullNestInfo& nestInfo)
{
const CompSlice* res = AllLoopReuseBlocking::SetBlocking(anal,nestInfo);
const CompSliceNest* curNest = nestInfo.GetNest();
#ifdef DEBUG
std::cerr << "Set blocking curNest: " << curNest->toString() << "\n";
#endif
unsigned size = curNest->NumberOfEntries();
#ifdef DEBUG
std::cerr << "size=" << size << "\n";
#endif
if (size <= 1 || !curNest->Entry(size-1)->SliceCommonLoop(curNest->Entry(0)))
{ /*QY: current loop nest is perfectly nested*/
const CompSliceDepGraphNode::NestInfo* nonperfect = DoNonPerfectBlocking(nestInfo);
if (nonperfect)
{ /*QY: extra loops inside; (not considered for blocking normally). */
const CompSliceNest* innerNest = nonperfect->GetNest();
assert(innerNest!=0); /*QY: this is why nonperfect is returned */
int reuseLevel = SliceNestReuseLevel(anal, *innerNest);
int j = 0, size = innerNest->NumberOfEntries();
for (; j < reuseLevel; ++j) blocksize.push_back(1);
for (; j < size; ++j) blocksize.push_back(GetDefaultBlockSize(innerNest->Entry(j)));
res = innerNest->Entry(size-1);
#ifdef DEBUG
std::cerr << "set nonperfect blocking\n";
#endif
}
}
return res;
}
const CompSlice* InnerLoopReuseBlocking ::
SetBlocking( CompSliceLocalityRegistry *anal,
const CompSliceDepGraphNode::FullNestInfo& nestInfo)
{
const CompSliceNest& n = *nestInfo.GetNest();
blocksize.clear();
unsigned num = n.NumberOfEntries();
int reuseLevel = SliceNestReuseLevel(anal, n);
blocksize.resize(reuseLevel + 1, 1);
for ( size_t index = reuseLevel+1; index < num; ++index)
blocksize.push_back(GetDefaultBlockSize(n[index]));
return n[num-1];
}
const CompSlice* OuterLoopReuseBlocking ::
SetBlocking( CompSliceLocalityRegistry *anal,
const CompSliceDepGraphNode::FullNestInfo& nestInfo)
{
const CompSliceNest& n = *nestInfo.GetNest();
blocksize.clear();
size_t num = n.NumberOfEntries();
int reuseLevel = SliceNestReuseLevel(anal, n);
for (int i = 0; i < reuseLevel; ++i)
blocksize.push_back(1);
size_t index;
for ( index = reuseLevel; index < num-spill; ++index) {
blocksize.push_back(GetDefaultBlockSize(n[index]));
}
for (; index < num; ++index)
blocksize.push_back(1);
return n[num-1];
}
