Skip to content

Commit

Permalink
Merge pull request The-OpenROAD-Project#4288 from The-OpenROAD-Projec…
Browse files Browse the repository at this point in the history 
…t-staging/mpl2-threads

Mpl2 threads
  • Loading branch information
@maliberty
maliberty authored Nov 21, 2023
2 parents 24cee04 + 86772c4 commit 28892de
Showing 1 changed file with 96 additions and 59 deletions.
155 changes: 96 additions & 59 deletions src/mpl2/src/hier_rtlmp.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3666,18 +3666,20 @@ void HierRTLMP::multiLevelMacroPlacement(Cluster* parent)
std::vector<SACoreSoftMacro*>
sa_containers; // store all the SA runs to avoid memory leakage
float best_cost = std::numeric_limits<float>::max();
// To give consistency across threads we check the solutions
// at a fixed interval independent of how many threads we are using.
const int check_interval = 10;
int begin_check = 0;
int end_check = std::min(check_interval, remaining_runs);
debugPrint(logger_,
MPL,
"hierarchical_macro_placement",
1,
"Start Simulated Annealing Core");
while (remaining_runs > 0) {
std::vector<SACoreSoftMacro*> sa_vector;
int run_thread
= (remaining_runs > num_threads_) ? num_threads_ : remaining_runs;
if (graphics_) {
run_thread = 1;
}
const int run_thread
= graphics_ ? 1 : std::min(remaining_runs, num_threads_);
for (int i = 0; i < run_thread; i++) {
debugPrint(logger_,
MPL,
Expand Down Expand Up @@ -3775,20 +3777,29 @@ void HierRTLMP::multiLevelMacroPlacement(Cluster* parent)
th.join();
}
}
remaining_runs -= run_thread;
// add macro tilings
for (auto& sa : sa_vector) {
sa_containers.push_back(sa); // add SA to containers
if (sa->isValid() && sa->getNormCost() < best_cost) {
best_cost = sa->getNormCost();
best_sa = sa;
sa_containers.push_back(sa);
}
while (sa_containers.size() >= end_check) {
while (begin_check < end_check) {
auto& sa = sa_containers[begin_check];
if (sa->isValid() && sa->getNormCost() < best_cost) {
best_cost = sa->getNormCost();
best_sa = sa;
}
++begin_check;
}
// add early stop mechanism
if (best_sa) {
break;
}
end_check = begin_check + std::min(check_interval, remaining_runs);
}
sa_vector.clear();
// add early stop mechanism
if (best_sa != nullptr) {
if (best_sa) {
break;
}
remaining_runs -= run_thread;
}
debugPrint(logger_,
MPL,
Expand Down Expand Up @@ -3903,18 +3914,17 @@ void HierRTLMP::multiLevelMacroPlacement(Cluster* parent)
best_sa = nullptr;
sa_containers.clear();
best_cost = std::numeric_limits<float>::max();
begin_check = 0;
end_check = std::min(check_interval, remaining_runs);
debugPrint(logger_,
MPL,
"hierarchical_macro_placement",
1,
"Start Simulated Annealing Core");
while (remaining_runs > 0) {
std::vector<SACoreSoftMacro*> sa_vector;
int run_thread
= (remaining_runs > num_threads_) ? num_threads_ : remaining_runs;
if (graphics_) {
run_thread = 1;
}
const int run_thread
= graphics_ ? 1 : std::min(remaining_runs, num_threads_);
for (int i = 0; i < run_thread; i++) {
debugPrint(logger_,
MPL,
Expand Down Expand Up @@ -4013,20 +4023,29 @@ void HierRTLMP::multiLevelMacroPlacement(Cluster* parent)
th.join();
}
}
remaining_runs -= run_thread;
// add macro tilings
for (auto& sa : sa_vector) {
sa_containers.push_back(sa); // add SA to containers
if (sa->isValid() && sa->getNormCost() < best_cost) {
best_cost = sa->getNormCost();
best_sa = sa;
sa_containers.push_back(sa);
}
while (sa_containers.size() >= end_check) {
while (begin_check < end_check) {
auto& sa = sa_containers[begin_check];
if (sa->isValid() && sa->getNormCost() < best_cost) {
best_cost = sa->getNormCost();
best_sa = sa;
}
++begin_check;
}
// add early stop mechanism
if (best_sa) {
break;
}
end_check = begin_check + std::min(check_interval, remaining_runs);
}
sa_vector.clear();
// add early stop mechanism
if (best_sa != nullptr) {
if (best_sa) {
break;
}
remaining_runs -= run_thread;
}
debugPrint(logger_,
MPL,
Expand Down Expand Up @@ -4443,8 +4462,12 @@ void HierRTLMP::multiLevelMacroPlacementWithoutBusPlanning(Cluster* parent)
int remaining_runs = target_util_list.size();
int run_id = 0;
SACoreSoftMacro* best_sa = nullptr;
std::vector<SACoreSoftMacro*>
sa_containers; // store all the SA runs to avoid memory leakage
std::vector<SACoreSoftMacro*> sa_containers;
// To give consistency across threads we check the solutions
// at a fixed interval independent of how many threads we are using.
const int check_interval = 10;
int begin_check = 0;
int end_check = std::min(check_interval, remaining_runs);
float best_cost = std::numeric_limits<float>::max();
debugPrint(logger_,
MPL,
Expand All @@ -4453,11 +4476,8 @@ void HierRTLMP::multiLevelMacroPlacementWithoutBusPlanning(Cluster* parent)
"Start Simulated Annealing Core");
while (remaining_runs > 0) {
std::vector<SACoreSoftMacro*> sa_vector;
int run_thread
= (remaining_runs > num_threads_) ? num_threads_ : remaining_runs;
if (graphics_) {
run_thread = 1;
}
const int run_thread
= graphics_ ? 1 : std::min(remaining_runs, num_threads_);
for (int i = 0; i < run_thread; i++) {
debugPrint(logger_,
MPL,
Expand Down Expand Up @@ -4555,20 +4575,29 @@ void HierRTLMP::multiLevelMacroPlacementWithoutBusPlanning(Cluster* parent)
th.join();
}
}
remaining_runs -= run_thread;
// add macro tilings
for (auto& sa : sa_vector) {
sa_containers.push_back(sa); // add SA to containers
if (sa->isValid() && sa->getNormCost() < best_cost) {
best_cost = sa->getNormCost();
best_sa = sa;
sa_containers.push_back(sa);
}
while (sa_containers.size() >= end_check) {
while (begin_check < end_check) {
auto& sa = sa_containers[begin_check];
if (sa->isValid() && sa->getNormCost() < best_cost) {
best_cost = sa->getNormCost();
best_sa = sa;
}
++begin_check;
}
// add early stop mechanism
if (best_sa) {
break;
}
end_check = begin_check + std::min(check_interval, remaining_runs);
}
sa_vector.clear();
// add early stop mechanism
if (best_sa != nullptr) {
if (best_sa) {
break;
}
remaining_runs -= run_thread;
}
debugPrint(logger_,
MPL,
Expand Down Expand Up @@ -4940,18 +4969,20 @@ void HierRTLMP::enhancedMacroPlacement(Cluster* parent)
std::vector<SACoreSoftMacro*>
sa_containers; // store all the SA runs to avoid memory leakage
float best_cost = std::numeric_limits<float>::max();
// To give consistency across threads we check the solutions
// at a fixed interval independent of how many threads we are using.
const int check_interval = 10;
int begin_check = 0;
int end_check = std::min(check_interval, remaining_runs);
debugPrint(logger_,
MPL,
"hierarchical_macro_placement",
1,
"Start Simulated Annealing Core");
while (remaining_runs > 0) {
std::vector<SACoreSoftMacro*> sa_vector;
int run_thread
= (remaining_runs > num_threads_) ? num_threads_ : remaining_runs;
if (graphics_) {
run_thread = 1;
}
const int run_thread
= graphics_ ? 1 : std::min(remaining_runs, num_threads_);
for (int i = 0; i < run_thread; i++) {
std::vector<SoftMacro> shaped_macros = macros; // copy for multithread
// determine the shape for each macro
Expand Down Expand Up @@ -5046,20 +5077,29 @@ void HierRTLMP::enhancedMacroPlacement(Cluster* parent)
th.join();
}
}
remaining_runs -= run_thread;
// add macro tilings
for (auto& sa : sa_vector) {
sa_containers.push_back(sa); // add SA to containers
if (sa->isValid() && sa->getNormCost() < best_cost) {
best_cost = sa->getNormCost();
best_sa = sa;
sa_containers.push_back(sa);
}
while (sa_containers.size() >= end_check) {
while (begin_check < end_check) {
auto& sa = sa_containers[begin_check];
if (sa->isValid() && sa->getNormCost() < best_cost) {
best_cost = sa->getNormCost();
best_sa = sa;
}
++begin_check;
}
// add early stop mechanism
if (best_sa) {
break;
}
end_check = begin_check + std::min(check_interval, remaining_runs);
}
sa_vector.clear();
// add early stop mechanism
if (best_sa != nullptr) {
if (best_sa) {
break;
}
remaining_runs -= run_thread;
}
debugPrint(logger_,
MPL,
Expand Down Expand Up @@ -5463,11 +5503,8 @@ void HierRTLMP::hardMacroClusterMacroPlacement(Cluster* cluster)
float best_cost = std::numeric_limits<float>::max();
while (remaining_runs > 0) {
std::vector<SACoreHardMacro*> sa_vector;
int run_thread
= (remaining_runs > num_threads_) ? num_threads_ : remaining_runs;
if (graphics_) {
run_thread = 1;
}
const int run_thread
= graphics_ ? 1 : std::min(remaining_runs, num_threads_);
for (int i = 0; i < run_thread; i++) {
// change the aspect ratio
const float width = outline_width * vary_factor_list[run_id++];
Expand Down

0 comments on commit 28892de

Please sign in to comment.