thesis version
This commit is contained in:
123
inc/alsk/executor/impl/dynamicpool.h
Normal file
123
inc/alsk/executor/impl/dynamicpool.h
Normal file
@ -0,0 +1,123 @@
|
||||
#ifndef ALSK_ALSK_EXECUTOR_IMPL_DYNAMICPOOL_H
|
||||
#define ALSK_ALSK_EXECUTOR_IMPL_DYNAMICPOOL_H
|
||||
|
||||
#include <algorithm>
|
||||
#include <future>
|
||||
#include <vector>
|
||||
|
||||
#include "../executorbase.h"
|
||||
#include "../executorstate.h"
|
||||
#include "../../skeleton/traits.h"
|
||||
|
||||
#include "../utility/pool.h"
|
||||
|
||||
namespace alsk {
|
||||
namespace exec {
|
||||
|
||||
template<typename S>
|
||||
struct DynamicPool: ExecutorBase {
|
||||
using Tag = alsk::tag::Parallel;
|
||||
|
||||
public:
|
||||
std::size_t maxTaskCount = 1'000;
|
||||
|
||||
public:
|
||||
template<typename Impl>
|
||||
void config(Impl& impl) {
|
||||
impl.state.executor.config(cores);
|
||||
}
|
||||
|
||||
template<typename Task, typename Impl, typename BTask, typename Parameters>
|
||||
void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
|
||||
std::size_t taskCount = std::min(maxTaskCount, n);
|
||||
|
||||
if(cores > 1 && taskCount > 1) {
|
||||
Info info;
|
||||
std::vector<std::future<void>> futures(taskCount);
|
||||
std::size_t const step = n/taskCount;
|
||||
std::size_t const remain = n - step*(taskCount-1);
|
||||
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
auto run = [&](std::size_t b, std::size_t k) {
|
||||
for(std::size_t i = 0; i < k; ++i)
|
||||
Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
|
||||
};
|
||||
|
||||
for(std::size_t i = 0; i < taskCount-1; ++i)
|
||||
futures[i] = state.executor.run([&, b=i*step, k=step]{ run(b, k); });
|
||||
futures[taskCount-1] = state.executor.run([&, b=(taskCount-1)*step, k=remain]{ run(b, k); });
|
||||
|
||||
state.executor.wait(futures);
|
||||
} else {
|
||||
Info info;
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
}
|
||||
}
|
||||
|
||||
template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
|
||||
Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
|
||||
std::size_t taskCount = std::min(maxTaskCount, n);
|
||||
|
||||
Value best{};
|
||||
|
||||
if(cores > 1 && taskCount > 1) {
|
||||
Info info;
|
||||
std::vector<std::future<void>> futures(taskCount);
|
||||
std::size_t const step = n/taskCount;
|
||||
std::size_t const remainBase = n - step*taskCount;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
auto run = [&](Value& out, std::size_t b, std::size_t k) {
|
||||
Value best{};
|
||||
|
||||
if(k)
|
||||
best = Task::execute(impl, task, b+0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < k; ++i) {
|
||||
Value current = Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, b+i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
|
||||
out = std::move(best);
|
||||
};
|
||||
|
||||
std::size_t start{};
|
||||
std::vector<Value> bests(taskCount);
|
||||
for(std::size_t i = 0; i < taskCount-1; ++i) {
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
futures[i] = state.executor.run([&, &best=bests[i], b=start, k=step+offset] { run(best, b, k); });
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
futures[taskCount-1] = state.executor.run([&, &best=bests[taskCount-1], b=start, k=step] { run(best, b, k); });
|
||||
|
||||
state.executor.wait(futures);
|
||||
|
||||
if(taskCount) best = std::move(bests[0]);
|
||||
for(std::size_t i = 1; i < taskCount; ++i)
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
|
||||
} else {
|
||||
Info info;
|
||||
if(n)
|
||||
best = Task::execute(impl, task, 0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < n; ++i) {
|
||||
Value current = Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
}
|
||||
|
||||
return best;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename S>
|
||||
struct ExecutorState<DynamicPool<S>>: util::Pool {};
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
195
inc/alsk/executor/impl/firstlevel/equi.h
Normal file
195
inc/alsk/executor/impl/firstlevel/equi.h
Normal file
@ -0,0 +1,195 @@
|
||||
#ifndef ALSK_ALSK_EXECUTOR_IMPL_FIRSTLEVEL_EQUI_H
|
||||
#define ALSK_ALSK_EXECUTOR_IMPL_FIRSTLEVEL_EQUI_H
|
||||
|
||||
#include <cmath>
|
||||
#include <thread>
|
||||
#include <set>
|
||||
#include <vector>
|
||||
|
||||
#include "../../executorbase.h"
|
||||
#include "../../executorstate.h"
|
||||
#include "../../../skeleton/traits.h"
|
||||
|
||||
namespace alsk {
|
||||
namespace exec {
|
||||
|
||||
template<typename S>
|
||||
struct FirstLevelEqui: ExecutorBase {
|
||||
using Tag = alsk::tag::Parallel;
|
||||
|
||||
public:
|
||||
struct Info {
|
||||
unsigned int parDepth;
|
||||
|
||||
Info(unsigned int parDepth = 0) noexcept: parDepth{parDepth} {}
|
||||
|
||||
Info par() const noexcept { return {parDepth+1}; }
|
||||
Info seq() const noexcept { return {parDepth}; }
|
||||
};
|
||||
|
||||
private:
|
||||
template<typename Impl>
|
||||
void buildSplit(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
|
||||
split.clear();
|
||||
|
||||
auto traverser = [](std::size_t, auto&& skl, auto&&... values) {
|
||||
using Skl = decltype(skl);
|
||||
using Traits = SkeletonTraitsT<std::decay_t<Skl>>;
|
||||
if(Traits::serial) return max(decltype(values)(values)...);
|
||||
return Traits::parallelizability(std::forward<Skl>(skl));
|
||||
};
|
||||
|
||||
auto firstLevelPar = SkeletonTraversal<S>::execute(impl.skeleton, traverser, 1ul);
|
||||
|
||||
split.insert(0);
|
||||
for(auto const& k: repeatability.coresList) {
|
||||
std::size_t start{};
|
||||
std::size_t const step = firstLevelPar/k;
|
||||
std::size_t remain = firstLevelPar - step*k;
|
||||
|
||||
for(unsigned int i = 0; i < k-1; ++i) {
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
start += step+offset;
|
||||
split.insert(start * (state.executor.parTasksCount/firstLevelPar));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
unsigned int threadLimit(Info const& info) const noexcept {
|
||||
return info.parDepth? 1 : cores;
|
||||
}
|
||||
|
||||
public:
|
||||
template<typename Impl>
|
||||
void config(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
state.executor.parTasksCount = impl.parallelTasksCount();;
|
||||
buildSplit(impl);
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextIdCount(Impl& impl, std::size_t) {
|
||||
typename Impl::State& state = impl.state;
|
||||
return state.executor.split.size();
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextId(Impl& impl, std::size_t id) { // O(log(n))
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
return std::distance(std::begin(split), split.upper_bound(id)) - 1;
|
||||
}
|
||||
|
||||
template<typename Task, typename Impl, typename BTask, typename Parameters>
|
||||
void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
|
||||
std::size_t const maxThreads = threadLimit(impl.executorInfo);
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
|
||||
if(nThreads > 1) {
|
||||
Info info = impl.executorInfo.par();
|
||||
std::vector<std::thread> threads(nThreads-1);
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
auto run = [&](std::size_t b, std::size_t k) {
|
||||
for(std::size_t i = 0; i < k; ++i)
|
||||
Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
|
||||
};
|
||||
|
||||
{
|
||||
std::size_t start{};
|
||||
for(std::size_t i = 0; i < nThreads-1; ++i) {
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
threads[i] = std::thread{run, start, step+offset};
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
run(start, step);
|
||||
}
|
||||
|
||||
for(std::thread& thread: threads) thread.join();
|
||||
} else {
|
||||
Info info = impl.executorInfo.seq();
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
}
|
||||
}
|
||||
|
||||
template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
|
||||
Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
|
||||
std::size_t const maxThreads = threadLimit(impl.executorInfo); // TODO? fix neighbours
|
||||
|
||||
Value best{};
|
||||
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
if(nThreads > 1) {
|
||||
Info info = impl.executorInfo.par();
|
||||
std::vector<std::thread> threads(nThreads-1);
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
auto run = [&](Value& out, std::size_t b, std::size_t k) {
|
||||
Value best{};
|
||||
|
||||
if(k)
|
||||
best = Task::execute(impl, task, b+0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < k; ++i) {
|
||||
Value current = Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, b+i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
|
||||
out = std::move(best);
|
||||
};
|
||||
|
||||
std::size_t start{};
|
||||
std::vector<Value> bests(nThreads);
|
||||
|
||||
{
|
||||
std::size_t i;
|
||||
for(i = 0; i < nThreads-1; ++i) {
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
threads[i] = std::thread{run, std::ref(bests[i]), start, step+offset};
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
run(bests[i], start, step);
|
||||
}
|
||||
|
||||
for(std::thread& thread: threads) thread.join();
|
||||
|
||||
if(nThreads) best = std::move(bests[0]);
|
||||
for(std::size_t i = 1; i < nThreads; ++i)
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
|
||||
} else {
|
||||
Info info = impl.executorInfo.seq();
|
||||
if(n)
|
||||
best = Task::execute(impl, task, 0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < n; ++i) {
|
||||
Value current = Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
}
|
||||
|
||||
return best;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename S>
|
||||
struct ExecutorState<FirstLevelEqui<S>> {
|
||||
std::size_t parTasksCount;
|
||||
std::set<std::size_t> split;
|
||||
};
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
168
inc/alsk/executor/impl/firstlevel/greedy.h
Normal file
168
inc/alsk/executor/impl/firstlevel/greedy.h
Normal file
@ -0,0 +1,168 @@
|
||||
#ifndef ALSK_ALSK_EXECUTOR_IMPL_FIRSTLEVEL_GREEDY_H
|
||||
#define ALSK_ALSK_EXECUTOR_IMPL_FIRSTLEVEL_GREEDY_H
|
||||
|
||||
#include <cmath>
|
||||
#include <thread>
|
||||
#include <set>
|
||||
#include <vector>
|
||||
|
||||
#include "../../executorbase.h"
|
||||
#include "../../executorstate.h"
|
||||
#include "../../../skeleton/traits.h"
|
||||
|
||||
namespace alsk {
|
||||
namespace exec {
|
||||
|
||||
template<typename S>
|
||||
struct FirstLevelGreedy: ExecutorBase {
|
||||
using Tag = alsk::tag::Parallel;
|
||||
|
||||
public:
|
||||
struct Info {
|
||||
unsigned int parDepth;
|
||||
};
|
||||
|
||||
private:
|
||||
template<typename Impl>
|
||||
void buildSplit(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
|
||||
split.clear();
|
||||
|
||||
auto traverser = [](std::size_t, auto&& skl, auto&&... values) {
|
||||
using Skl = decltype(skl);
|
||||
using Traits = alsk::SkeletonTraitsT<Skl>;
|
||||
if(Traits::serial) return max(values...);
|
||||
return Traits::parallelizability(std::forward<Skl>(skl));
|
||||
};
|
||||
|
||||
auto firstLevelPar = SkeletonTraversal<S>::execute(impl.skeleton, traverser, 1ul);
|
||||
|
||||
split.insert(0);
|
||||
for(auto const& k: repeatability.coresList) {
|
||||
std::size_t start{};
|
||||
std::size_t const step = (firstLevelPar + k-1)/k;
|
||||
std::size_t const rk = (firstLevelPar + step-1)/step;
|
||||
|
||||
for(unsigned int i = 0; i < rk; ++i, start += step)
|
||||
split.insert(start * (state.executor.parTasksCount/firstLevelPar));
|
||||
}
|
||||
}
|
||||
|
||||
unsigned int threadLimit(unsigned int level) const { return level? 1 : cores; }
|
||||
|
||||
public:
|
||||
template<typename Impl>
|
||||
void config(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
state.executor.parTasksCount = impl.parallelTasksCount();;
|
||||
buildSplit(impl);
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextIdCount(Impl& impl, std::size_t) {
|
||||
typename Impl::State& state = impl.state;
|
||||
return state.executor.split.size();
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextId(Impl& impl, std::size_t id) { // O(log(n))
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
return std::distance(std::begin(split), split.upper_bound(id)) - 1;
|
||||
}
|
||||
|
||||
template<typename Task, typename Impl, typename BTask, typename Parameters>
|
||||
void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
|
||||
auto const& parDepth = impl.executorInfo.parDepth;
|
||||
std::size_t const maxThreads = threadLimit(parDepth);
|
||||
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
if(nThreads > 1) {
|
||||
Info info{parDepth+1};
|
||||
std::vector<std::thread> threads(nThreads-1);
|
||||
std::size_t const step = std::round(static_cast<double>(n)/nThreads);
|
||||
|
||||
auto run = [&](std::size_t b, std::size_t k) {
|
||||
for(std::size_t i = 0; i < k; ++i)
|
||||
Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
|
||||
};
|
||||
|
||||
for(std::size_t i = 0; i < nThreads-1; ++i)
|
||||
threads[i] = std::thread{run, i*step, step};
|
||||
|
||||
run((nThreads-1)*step, n-(nThreads-1)*step);
|
||||
|
||||
for(std::thread& thread: threads) thread.join();
|
||||
} else {
|
||||
Info info{parDepth};
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
}
|
||||
}
|
||||
|
||||
template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
|
||||
Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
|
||||
auto const& parDepth = impl.executorInfo.parDepth;
|
||||
std::size_t const maxThreads = threadLimit(parDepth); // TODO fix neighbours
|
||||
|
||||
Value best{};
|
||||
|
||||
std::size_t const nThreadsBase = std::min(n, maxThreads);
|
||||
if(nThreadsBase > 1) {
|
||||
Info info{parDepth+1};
|
||||
std::size_t const step = (n+nThreadsBase-1)/nThreadsBase;
|
||||
std::size_t const nThreads = (n+step-1)/step;
|
||||
std::vector<std::thread> threads(nThreads-1);
|
||||
|
||||
auto run = [&](Value& out, std::size_t b, std::size_t k) {
|
||||
Value best{};
|
||||
|
||||
if(k)
|
||||
best = Task::execute(impl, task, b+0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < k; ++i) {
|
||||
Value current = Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, b+i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
|
||||
out = std::move(best);
|
||||
};
|
||||
|
||||
std::size_t start{};
|
||||
std::vector<Value> bests(nThreads);
|
||||
for(std::size_t i = 0; i < nThreads-1; ++i, start += step)
|
||||
threads[i] = std::thread{run, std::ref(bests[i]), start, step};
|
||||
|
||||
run(bests[nThreads-1], start, n - step*(nThreads-1));
|
||||
|
||||
for(std::thread& thread: threads) thread.join();
|
||||
|
||||
if(nThreads) best = std::move(bests[0]);
|
||||
for(std::size_t i = 1; i < nThreads; ++i)
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
|
||||
} else {
|
||||
Info info{parDepth};
|
||||
if(n)
|
||||
best = Task::execute(impl, task, 0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < n; ++i) {
|
||||
Value current = Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
}
|
||||
|
||||
return best;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename S>
|
||||
struct ExecutorState<FirstLevelGreedy<S>> {
|
||||
std::size_t parTasksCount;
|
||||
std::set<std::size_t> split;
|
||||
};
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
129
inc/alsk/executor/impl/firstlevel/noopti.h
Normal file
129
inc/alsk/executor/impl/firstlevel/noopti.h
Normal file
@ -0,0 +1,129 @@
|
||||
#ifndef ALSK_ALSK_EXECUTOR_IMPL_FIRSTLEVEL_NOOPTI_H
|
||||
#define ALSK_ALSK_EXECUTOR_IMPL_FIRSTLEVEL_NOOPTI_H
|
||||
|
||||
#include <thread>
|
||||
#include <set>
|
||||
#include <cmath>
|
||||
#include <vector>
|
||||
|
||||
#include "../../executorbase.h"
|
||||
#include "../../executorstate.h"
|
||||
#include "../../../skeleton/traits.h"
|
||||
|
||||
namespace alsk {
|
||||
namespace exec {
|
||||
|
||||
template<typename S>
|
||||
struct FirstLevelNoOpti: ExecutorBase {
|
||||
using Tag = alsk::tag::Parallel;
|
||||
|
||||
public:
|
||||
struct Info {
|
||||
unsigned int parDepth;
|
||||
};
|
||||
|
||||
private:
|
||||
unsigned int threadLimit(unsigned int level) const { return level? 1 : cores; }
|
||||
|
||||
public:
|
||||
template<typename Impl>
|
||||
std::size_t contextIdCount(Impl&, std::size_t count) { return count; }
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextId(Impl&, std::size_t id) { return id; }
|
||||
|
||||
template<typename Task, typename Impl, typename BTask, typename Parameters>
|
||||
void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
|
||||
auto const& parDepth = impl.executorInfo.parDepth;
|
||||
std::size_t const maxThreads = threadLimit(parDepth);
|
||||
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
if(nThreads > 1) {
|
||||
Info info{parDepth+1};
|
||||
std::vector<std::thread> threads(nThreads-1);
|
||||
std::size_t const step = std::round(static_cast<double>(n)/nThreads);
|
||||
|
||||
auto run = [&](std::size_t b, std::size_t k) {
|
||||
for(std::size_t i = 0; i < k; ++i)
|
||||
Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
|
||||
};
|
||||
|
||||
for(std::size_t i = 0; i < nThreads-1; ++i)
|
||||
threads[i] = std::thread{run, i*step, step};
|
||||
|
||||
run((nThreads-1)*step, n-(nThreads-1)*step);
|
||||
|
||||
for(std::thread& thread: threads) thread.join();
|
||||
|
||||
} else {
|
||||
Info info{parDepth};
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
}
|
||||
}
|
||||
|
||||
template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
|
||||
Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
|
||||
auto const& parDepth = impl.executorInfo.parDepth;
|
||||
std::size_t const maxThreads = threadLimit(parDepth); // TODO fix neighbours
|
||||
|
||||
Value best{};
|
||||
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
if(nThreads > 1) {
|
||||
Info info{parDepth+1};
|
||||
std::vector<std::thread> threads(nThreads-1);
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
auto run = [&](Value& out, std::size_t b, std::size_t k) {
|
||||
Value best{};
|
||||
|
||||
if(k)
|
||||
best = Task::execute(impl, task, b+0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < k; ++i) {
|
||||
Value current = Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, b+i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
|
||||
out = std::move(best);
|
||||
};
|
||||
|
||||
std::size_t start{};
|
||||
std::vector<Value> bests(nThreads);
|
||||
for(std::size_t i = 0; i < nThreads-1; ++i) {
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
threads[i] = std::thread{run, std::ref(bests[i]), start, step+offset};
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
run(bests[nThreads-1], start, step);
|
||||
|
||||
for(std::thread& thread: threads) thread.join();
|
||||
|
||||
if(nThreads) best = std::move(bests[0]);
|
||||
for(std::size_t i = 1; i < nThreads; ++i)
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
|
||||
} else {
|
||||
Info info{parDepth};
|
||||
if(n)
|
||||
best = Task::execute(impl, task, 0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < n; ++i) {
|
||||
Value current = Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
}
|
||||
|
||||
return best;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename S>
|
||||
struct ExecutorState<FirstLevelNoOpti<S>> {};
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
55
inc/alsk/executor/impl/sequential.h
Normal file
55
inc/alsk/executor/impl/sequential.h
Normal file
@ -0,0 +1,55 @@
|
||||
#ifndef ALSK_ALSK_EXECUTOR_IMPL_SEQUENTIAL_H
|
||||
#define ALSK_ALSK_EXECUTOR_IMPL_SEQUENTIAL_H
|
||||
|
||||
#include <set>
|
||||
#include <cmath>
|
||||
#include <vector>
|
||||
|
||||
#include "../executorbase.h"
|
||||
#include "../executorstate.h"
|
||||
#include "../../skeleton/traits.h"
|
||||
|
||||
namespace alsk {
|
||||
namespace exec {
|
||||
|
||||
template<typename S>
|
||||
struct Sequential: ExecutorBase {
|
||||
using Tag = alsk::tag::Sequential;
|
||||
|
||||
public:
|
||||
template<typename Impl>
|
||||
std::size_t contextIdCount(Impl&, std::size_t) { return 1; }
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextId(Impl&, std::size_t) { return 0; }
|
||||
|
||||
template<typename Task, typename Impl, typename BTask, typename Parameters>
|
||||
void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
|
||||
Info info;
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
}
|
||||
|
||||
template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
|
||||
Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
|
||||
Info info;
|
||||
Value best{};
|
||||
|
||||
if(n)
|
||||
best = Task::execute(impl, task, 0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < n; ++i) {
|
||||
Value current = Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
|
||||
return best;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename S>
|
||||
struct ExecutorState<Sequential<S>> {};
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
272
inc/alsk/executor/impl/staticpool.h
Normal file
272
inc/alsk/executor/impl/staticpool.h
Normal file
@ -0,0 +1,272 @@
|
||||
#ifndef ALSK_ALSK_EXECUTOR_IMPL_STATICPOOL_H
|
||||
#define ALSK_ALSK_EXECUTOR_IMPL_STATICPOOL_H
|
||||
|
||||
#include <algorithm>
|
||||
#include <cmath>
|
||||
#include <future>
|
||||
#include <map>
|
||||
#include <set>
|
||||
#include <vector>
|
||||
|
||||
#include "../executorbase.h"
|
||||
#include "../executorstate.h"
|
||||
#include "../../skeleton/traits.h"
|
||||
|
||||
#include "../utility/staticpool.h"
|
||||
|
||||
namespace alsk {
|
||||
namespace exec {
|
||||
|
||||
template<typename S>
|
||||
struct StaticPool: ExecutorBase {
|
||||
using Tag = alsk::tag::Parallel;
|
||||
|
||||
public:
|
||||
struct Info {
|
||||
std::size_t cores;
|
||||
std::size_t offset;
|
||||
|
||||
Info(std::size_t cores = 0, std::size_t offset = 0):
|
||||
cores{cores}, offset{offset} {}
|
||||
};
|
||||
|
||||
private:
|
||||
auto buildSplitFor(S& s, std::size_t cores) {
|
||||
std::map<std::size_t, std::size_t> ms;
|
||||
|
||||
auto populateSplitImpl = [&ms, totalCores=cores](
|
||||
auto& buildSplitImpl, auto& s,
|
||||
std::size_t maxThreads, std::size_t thOffset,
|
||||
std::size_t n, std::size_t idStep, std::size_t id, bool isRemainder
|
||||
) {
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
if(nThreads > 0) {
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
std::size_t const coresA = maxThreads/nThreads;
|
||||
std::size_t const coresB = remainBase? maxThreads/remainBase : 1;
|
||||
|
||||
std::size_t start = 0;
|
||||
for(std::size_t i = 0; i < nThreads; ++i) {
|
||||
std::size_t thNum = thOffset + i*coresA;
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
|
||||
if(!ms.count(id+start*idStep))
|
||||
ms[id+start*idStep] = thNum;
|
||||
|
||||
for(std::size_t j = 0; j < step; ++j)
|
||||
buildSplitImpl(s, coresA, thNum, id+(start+j)*idStep, false);
|
||||
if(offset)
|
||||
buildSplitImpl(s, coresB, thNum, id+(start+step)*idStep, true);
|
||||
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
if(isRemainder) ms[id+start*idStep] = totalCores;
|
||||
} else {
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
buildSplitImpl(s, maxThreads, thOffset, id+i*idStep, false);
|
||||
}
|
||||
};
|
||||
|
||||
auto buildSplitImpl = makeRecursiveLambda(
|
||||
[&populateSplitImpl](
|
||||
auto buildSplitImpl,
|
||||
auto& s, auto maxThreads, auto thOffset,
|
||||
auto id, bool isRemainder
|
||||
) {
|
||||
auto idStep = skeletonStep(s);
|
||||
auto populateSplit = [&](auto& s, std::size_t n) {
|
||||
if(!idStep) return;
|
||||
populateSplitImpl(buildSplitImpl, s, maxThreads, thOffset, n, idStep, id, isRemainder);
|
||||
};
|
||||
|
||||
|
||||
skeletonTraversal(s, populateSplit);
|
||||
}
|
||||
);
|
||||
|
||||
buildSplitImpl(s, cores, 0ul, 0ul, false);
|
||||
|
||||
return ms;
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
void buildSplit(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
|
||||
split.clear();
|
||||
split.insert(0);
|
||||
|
||||
for(auto cores: repeatability.coresList) {
|
||||
std::size_t curThread = 0;
|
||||
for(auto p: buildSplitFor(impl.skeleton, cores)) { // TODO: C++17
|
||||
if(std::get<1>(p) != curThread) {
|
||||
curThread = std::get<1>(p);
|
||||
split.insert(std::get<0>(p));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
std::size_t threadLimit(Info const& info) const {
|
||||
auto const& lCores = info.cores;
|
||||
return lCores? lCores : cores;
|
||||
}
|
||||
|
||||
public:
|
||||
template<typename Impl>
|
||||
void config(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
impl.executorInfo.cores = cores;
|
||||
impl.executorInfo.offset = 0;
|
||||
state.executor.config(cores);
|
||||
|
||||
state.executor.parTasksCount = impl.parallelTasksCount();;
|
||||
buildSplit(impl);
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextIdCount(Impl& impl, std::size_t) {
|
||||
typename Impl::State& state = impl.state;
|
||||
return state.executor.split.size();
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextId(Impl& impl, std::size_t id) { // O(log(n))
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
return std::distance(std::begin(split), split.upper_bound(id)) - 1;
|
||||
}
|
||||
|
||||
template<typename Task, typename Impl, typename BTask, typename Parameters>
|
||||
void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
|
||||
std::size_t const maxThreads = threadLimit(impl.executorInfo);
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
|
||||
if(nThreads > 0) {
|
||||
std::vector<std::future<void>> futures(nThreads);
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
std::size_t const coresA = maxThreads/nThreads; // cores for sub tasks in main cases
|
||||
std::size_t const coresB = remainBase? maxThreads/remainBase : 1; // cores for remaining tasks
|
||||
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
auto run = [&](std::size_t b, std::size_t k, bool offset, std::size_t thOffset) {
|
||||
Info infoA{coresA, thOffset}, infoB{coresB, thOffset};
|
||||
|
||||
std::size_t i;
|
||||
for(i = 0; i < k; ++i)
|
||||
Task::execute(impl, task, b+i, infoA, parameters, std::tuple<>{});
|
||||
if(offset)
|
||||
Task::execute(impl, task, b+i, infoB, parameters, std::tuple<>{});
|
||||
};
|
||||
|
||||
for(std::size_t i = 0, start = 0; i < nThreads; ++i) {
|
||||
std::size_t thNum = impl.executorInfo.offset + i*coresA;
|
||||
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
|
||||
auto task = [&run, start, step, offset, thNum]{ run(start, step, offset, thNum); };
|
||||
futures[i] = state.executor.run(thNum, std::move(task));
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
state.executor.wait(futures);
|
||||
} else {
|
||||
Info info{impl.executorInfo};
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
}
|
||||
}
|
||||
|
||||
template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
|
||||
Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
|
||||
std::size_t const maxThreads = threadLimit(impl.executorInfo);
|
||||
|
||||
Value best{};
|
||||
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
if(nThreads > 0) {
|
||||
std::vector<std::future<void>> futures(nThreads);
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
std::size_t const coresA = maxThreads/nThreads; // cores for sub tasks in main cases
|
||||
std::size_t const coresB = remainBase? maxThreads/remainBase : 1; // cores for remaining tasks
|
||||
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
auto run = [&](Value& out, std::size_t b, std::size_t k, bool offset, std::size_t thOffset) {
|
||||
Value best{};
|
||||
Info infoA{coresA, thOffset}, infoB{coresB, thOffset};
|
||||
|
||||
if(k) {
|
||||
best = Task::execute(impl, task, b+0, infoA, parameters, std::tuple<>{});
|
||||
|
||||
std::size_t i;
|
||||
for(i = 1; i < k; ++i) {
|
||||
Value current = Task::execute(impl, task, b+i, infoA, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, b+i, infoA, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
|
||||
if(offset) {
|
||||
Value current = Task::execute(impl, task, b+i, infoB, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, b+i, infoB, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
}
|
||||
|
||||
out = std::move(best);
|
||||
};
|
||||
|
||||
std::vector<Value> bests(nThreads);
|
||||
for(std::size_t i = 0, start = 0; i < nThreads; ++i) {
|
||||
std::size_t thNum = impl.executorInfo.offset + i*coresA;
|
||||
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
|
||||
auto task = [&, &best=bests[i], start, step, offset, thNum]{ run(best, start, step, offset, thNum); };
|
||||
futures[i] = state.executor.run(thNum, std::move(task));
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
state.executor.wait(futures);
|
||||
|
||||
if(nThreads) best = std::move(bests[0]);
|
||||
for(std::size_t i = 1; i < nThreads; ++i)
|
||||
best = Select::execute(impl, select, i, impl.executorInfo, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
|
||||
} else {
|
||||
Info info{impl.executorInfo};
|
||||
if(n)
|
||||
best = Task::execute(impl, task, 0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < n; ++i) {
|
||||
Value current = Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
}
|
||||
|
||||
return best;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename S>
|
||||
struct ExecutorState<StaticPool<S>>: util::StaticPool {
|
||||
std::size_t parTasksCount;
|
||||
std::set<std::size_t> split;
|
||||
};
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
110
inc/alsk/executor/impl/staticpoolid.h
Normal file
110
inc/alsk/executor/impl/staticpoolid.h
Normal file
@ -0,0 +1,110 @@
|
||||
#ifndef ALSK_ALSK_EXECUTOR_IMPL_STATICPOOLID_H
|
||||
#define ALSK_ALSK_EXECUTOR_IMPL_STATICPOOLID_H
|
||||
|
||||
#include <algorithm>
|
||||
#include <cmath>
|
||||
#include <future>
|
||||
#include <list>
|
||||
#include <set>
|
||||
#include <vector>
|
||||
|
||||
#include "../executorbase.h"
|
||||
#include "../executorstate.h"
|
||||
#include "../../skeleton/traits.h"
|
||||
|
||||
#include "../utility/staticpool.h"
|
||||
|
||||
namespace alsk {
|
||||
namespace exec {
|
||||
|
||||
template<typename S>
|
||||
struct StaticPoolId: ExecutorBase {
|
||||
using Tag = alsk::tag::Parallel;
|
||||
|
||||
private:
|
||||
template<typename Impl>
|
||||
void buildSplit(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
|
||||
split.clear();
|
||||
split.insert(0);
|
||||
|
||||
auto const n = static_cast<double>(state.executor.upperId);
|
||||
for(auto cores: repeatability.coresList)
|
||||
for(std::size_t i = 1; i < cores; ++i)
|
||||
split.insert(std::ceil(n/cores * i));
|
||||
}
|
||||
|
||||
public:
|
||||
template<typename Impl>
|
||||
void config(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
state.executor.config(cores);
|
||||
state.executor.upperId = impl.parallelTasksCount();
|
||||
buildSplit(impl);
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextIdCount(Impl& impl, std::size_t) {
|
||||
typename Impl::State& state = impl.state;
|
||||
return state.executor.split.size();
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextId(Impl& impl, std::size_t id) { // O(log(n))
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
return std::distance(std::begin(split), split.upper_bound(id)) - 1;
|
||||
}
|
||||
|
||||
template<typename Task, typename Impl, typename BTask, typename Parameters>
|
||||
void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
|
||||
std::list<std::future<void>> futures;
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
for(std::size_t i = 0; i < n; ++i) {
|
||||
std::size_t thNum = cores * (impl.id + impl.skeleton.step * i) / state.executor.upperId;
|
||||
|
||||
auto thTask = [&, i]{ Task::execute(impl, task, i, Info{}, parameters, std::tuple<>{}); };
|
||||
futures.emplace_back(state.executor.run(thNum, std::move(thTask)));
|
||||
}
|
||||
|
||||
state.executor.wait(futures);
|
||||
}
|
||||
|
||||
template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
|
||||
Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
|
||||
Value best{};
|
||||
std::vector<Value> bests(n);
|
||||
|
||||
std::list<std::future<void>> futures;
|
||||
typename Impl::State& state = impl.state;
|
||||
|
||||
for(std::size_t i = 0; i < n; ++i) {
|
||||
std::size_t thNum = cores * (impl.id + impl.skeleton.step * i) / state.executor.upperId;
|
||||
|
||||
auto thTask = [&, &best = bests[i], i]{ best = Task::execute(impl, task, i, Info{}, parameters, std::tuple<>{}); };
|
||||
futures.emplace_back(state.executor.run(thNum, std::move(thTask)));
|
||||
}
|
||||
|
||||
state.executor.wait(futures);
|
||||
|
||||
if(n) best = std::move(bests[0]);
|
||||
for(std::size_t i = 1; i < n; ++i)
|
||||
best = Select::execute(impl, select, i, impl.executorInfo, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
|
||||
|
||||
return best;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename S>
|
||||
struct ExecutorState<StaticPoolId<S>>: util::StaticPool {
|
||||
std::size_t upperId;
|
||||
std::set<std::size_t> split;
|
||||
};
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
268
inc/alsk/executor/impl/staticthread.h
Normal file
268
inc/alsk/executor/impl/staticthread.h
Normal file
@ -0,0 +1,268 @@
|
||||
#ifndef ALSK_ALSK_EXECUTOR_IMPL_STATICTHREAD_H
|
||||
#define ALSK_ALSK_EXECUTOR_IMPL_STATICTHREAD_H
|
||||
|
||||
#include <algorithm>
|
||||
#include <cmath>
|
||||
#include <map>
|
||||
#include <set>
|
||||
#include <thread>
|
||||
#include <vector>
|
||||
|
||||
#include "../executorbase.h"
|
||||
#include "../executorstate.h"
|
||||
#include "../../skeleton/traits.h"
|
||||
|
||||
namespace alsk {
|
||||
namespace exec {
|
||||
|
||||
template<typename S>
|
||||
struct StaticThread: ExecutorBase {
|
||||
using Tag = alsk::tag::Parallel;
|
||||
|
||||
public:
|
||||
struct Info {
|
||||
std::size_t cores;
|
||||
|
||||
Info(std::size_t cores = 0):
|
||||
cores{cores} {}
|
||||
};
|
||||
|
||||
private:
|
||||
auto buildSplitFor(S& s, std::size_t cores) {
|
||||
std::map<std::size_t, std::size_t> ms;
|
||||
|
||||
auto populateSplitImpl = [&ms, totalCores=cores](
|
||||
auto& buildSplitImpl, auto& s,
|
||||
std::size_t maxThreads, std::size_t thOffset,
|
||||
std::size_t n, std::size_t idStep, std::size_t id, bool isRemainder
|
||||
) {
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
if(nThreads > 0) {
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
std::size_t const coresA = maxThreads/nThreads;
|
||||
std::size_t const coresB = remainBase? maxThreads/remainBase : 1;
|
||||
|
||||
std::size_t start = 0;
|
||||
for(std::size_t i = 0; i < nThreads; ++i) {
|
||||
std::size_t thNum = thOffset + i*coresA;
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
|
||||
if(!ms.count(id+start*idStep))
|
||||
ms[id+start*idStep] = thNum;
|
||||
|
||||
for(std::size_t j = 0; j < step; ++j)
|
||||
buildSplitImpl(s, coresA, thNum, id+(start+j)*idStep, false);
|
||||
if(offset)
|
||||
buildSplitImpl(s, coresB, thNum, id+(start+step)*idStep, true);
|
||||
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
if(isRemainder) ms[id+start*idStep] = totalCores;
|
||||
} else {
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
buildSplitImpl(s, maxThreads, thOffset, id+i*idStep, false);
|
||||
}
|
||||
};
|
||||
|
||||
auto buildSplitImpl = makeRecursiveLambda(
|
||||
[&populateSplitImpl](
|
||||
auto buildSplitImpl,
|
||||
auto& s, auto maxThreads, auto thOffset,
|
||||
auto id, bool isRemainder
|
||||
) {
|
||||
auto idStep = skeletonStep(s);
|
||||
auto populateSplit = [&](auto& s, std::size_t n) {
|
||||
if(!idStep) return;
|
||||
populateSplitImpl(buildSplitImpl, s, maxThreads, thOffset, n, idStep, id, isRemainder);
|
||||
};
|
||||
|
||||
|
||||
skeletonTraversal(s, populateSplit);
|
||||
}
|
||||
);
|
||||
|
||||
buildSplitImpl(s, cores, 0ul, 0ul, false);
|
||||
|
||||
return ms;
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
void buildSplit(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
|
||||
split.clear();
|
||||
split.insert(0);
|
||||
|
||||
for(auto cores: repeatability.coresList) {
|
||||
std::size_t curThread = 0;
|
||||
for(auto p: buildSplitFor(impl.skeleton, cores)) { // TODO: C++17
|
||||
if(std::get<1>(p) != curThread) {
|
||||
curThread = std::get<1>(p);
|
||||
split.insert(std::get<0>(p));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
std::size_t threadLimit(Info const& info) const {
|
||||
auto const& lCores = info.cores;
|
||||
return lCores? lCores : cores;
|
||||
}
|
||||
|
||||
public:
|
||||
template<typename Impl>
|
||||
void config(Impl& impl) {
|
||||
typename Impl::State& state = impl.state;
|
||||
impl.executorInfo.cores = cores;
|
||||
|
||||
state.executor.parTasksCount = impl.parallelTasksCount();;
|
||||
buildSplit(impl);
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextIdCount(Impl& impl, std::size_t) {
|
||||
typename Impl::State& state = impl.state;
|
||||
return state.executor.split.size();
|
||||
}
|
||||
|
||||
template<typename Impl>
|
||||
std::size_t contextId(Impl& impl, std::size_t id) { // O(log(n))
|
||||
typename Impl::State& state = impl.state;
|
||||
auto& split = state.executor.split;
|
||||
return std::distance(std::begin(split), split.upper_bound(id)) - 1;
|
||||
}
|
||||
|
||||
template<typename Task, typename Impl, typename BTask, typename Parameters>
|
||||
void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
|
||||
std::size_t const maxThreads = threadLimit(impl.executorInfo);
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
|
||||
if(nThreads > 1) {
|
||||
std::vector<std::thread> threads(nThreads-1);
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
std::size_t const coresA = maxThreads/nThreads; // cores for sub tasks in main cases
|
||||
std::size_t const coresB = remainBase? maxThreads/remainBase : 1; // cores for remaining tasks
|
||||
|
||||
auto run = [&](std::size_t b, std::size_t k, bool offset = false) {
|
||||
Info infoA{coresA}, infoB{coresB};
|
||||
|
||||
std::size_t i;
|
||||
for(i = 0; i < k; ++i)
|
||||
Task::execute(impl, task, b+i, infoA, parameters, std::tuple<>{});
|
||||
if(offset)
|
||||
Task::execute(impl, task, b+i, infoB, parameters, std::tuple<>{});
|
||||
};
|
||||
|
||||
{
|
||||
std::size_t start = 0;
|
||||
for(std::size_t i = 0; i < nThreads-1; ++i) {
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
|
||||
auto task = [&run, start, step, offset]{ run(start, step, offset); };
|
||||
threads[i] = std::thread{std::move(task)};
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
run(start, step);
|
||||
}
|
||||
|
||||
for(std::thread& thread: threads) thread.join();
|
||||
} else {
|
||||
Info info{impl.executorInfo};
|
||||
for(std::size_t i = 0; i < n; ++i)
|
||||
Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
}
|
||||
}
|
||||
|
||||
template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
|
||||
Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
|
||||
std::size_t const maxThreads = threadLimit(impl.executorInfo);
|
||||
|
||||
Value best{};
|
||||
|
||||
std::size_t const nThreads = std::min(n, maxThreads);
|
||||
if(nThreads > 1) {
|
||||
std::vector<std::thread> threads(nThreads-1);
|
||||
std::size_t const step = n/nThreads;
|
||||
std::size_t const remainBase = n - step*nThreads;
|
||||
std::size_t remain = remainBase;
|
||||
|
||||
std::size_t const coresA = maxThreads/nThreads; // cores for sub tasks in main cases
|
||||
std::size_t const coresB = remainBase? maxThreads/remainBase : 1; // cores for remaining tasks
|
||||
|
||||
auto run = [&](Value& out, std::size_t b, std::size_t k, bool offset = false) {
|
||||
Value best{};
|
||||
Info infoA{coresA}, infoB{coresB};
|
||||
|
||||
if(k) {
|
||||
best = Task::execute(impl, task, b+0, infoA, parameters, std::tuple<>{});
|
||||
|
||||
std::size_t i;
|
||||
for(i = 1; i < k; ++i) {
|
||||
Value current = Task::execute(impl, task, b+i, infoA, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, b+i, infoA, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
|
||||
if(offset) {
|
||||
Value current = Task::execute(impl, task, b+i, infoB, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, b+i, infoB, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
}
|
||||
|
||||
out = std::move(best);
|
||||
};
|
||||
|
||||
std::vector<Value> bests(nThreads);
|
||||
{
|
||||
std::size_t start = 0;
|
||||
for(std::size_t i = 0; i < nThreads-1; ++i) {
|
||||
std::size_t offset = !!remain;
|
||||
remain -= offset;
|
||||
|
||||
auto task = [&, &best=bests[i], start, step, offset]{ run(best, start, step, offset); };
|
||||
threads[i] = std::thread{std::move(task)};
|
||||
start += step+offset;
|
||||
}
|
||||
|
||||
run(bests[nThreads-1], start, step);
|
||||
}
|
||||
|
||||
for(std::thread& thread: threads) thread.join();
|
||||
|
||||
if(nThreads) best = std::move(bests[0]);
|
||||
for(std::size_t i = 1; i < nThreads; ++i)
|
||||
best = Select::execute(impl, select, i, impl.executorInfo, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
|
||||
} else {
|
||||
Info info{impl.executorInfo};
|
||||
if(n)
|
||||
best = Task::execute(impl, task, 0, info, parameters, std::tuple<>{});
|
||||
for(std::size_t i = 1; i < n; ++i) {
|
||||
Value current = Task::execute(impl, task, i, info, parameters, std::tuple<>{});
|
||||
best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
|
||||
}
|
||||
}
|
||||
|
||||
return best;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename S>
|
||||
struct ExecutorState<StaticThread<S>> {
|
||||
std::size_t parTasksCount;
|
||||
std::set<std::size_t> split;
|
||||
};
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
Reference in New Issue
Block a user