Light, self-contained, thread pool-based implementation of C++17 parallel standard library algorithms.
C++17 introduced parallel overloads of standard library algorithms that accept an Execution Policy as the first argument. Policies specify limits on how the implementation may parallelize the algorithm, enabling methods like threads, vectorization, or even GPU. Policies can be supplied by the compiler or by libraries like this one.
std::sort(std::execution::par, vec.begin(), vec.end());
// ^^^^^^^^^^^^^^^^^^^ native C++17 parallel Execution Policy Unfortunately compiler support varies. Quick summary of compilers' default standard libraries:
| Linux | macOS | Windows | |
|---|---|---|---|
| GCC 9+ | TBB Required | TBB Required | TBB Required |
| GCC 8- | ❌ | ❌ | ❌ |
| Clang (libc++) | ❌ | ❌ | ❌ |
| Clang (libstdc++) | TBB Required | TBB Required | TBB Required |
| Apple Clang | ❌ | ||
| MSVC 15.7+ (2017) | âś… | ||
| Parallel STL | TBB Required | TBB Required | TBB Required |
| poolSTL | âś…* | âś…* | âś…* |
PoolSTL is a supplement to fill in the support gaps. It is not a full implementation; only the basics are covered. However, it is small, easy to integrate, and has no external dependencies. A good backup to the other options.
Use poolSTL exclusively, or only on platforms lacking native support, or only if TBB is not present.
Supports C++11 and higher. Algorithms introduced in C++17 require C++17 or higher.
Tested in CI on GCC 7+, Clang/LLVM 5+, Apple Clang, MSVC, MinGW, and Emscripten.
Algorithms are added on an as-needed basis. If you need one open an issue or contribute a PR.
Limitations: All iterators must be random access. No nested parallel calls.
all_of,any_of,none_ofcopy,copy_ncount,count_iffill,fill_nfind,find_if,find_if_notfor_each,for_each_npartitionsort,stable_sorttransform
exclusive_scan(C++17 only)reducetransform_reduce(C++17 only)
All in std:: namespace.
poolstl::iota_iter- Iterate over integers. Same as iterating over output ofstd::iotabut without materializing anything. Iterator version ofstd::ranges::iota_view.poolstl::for_each_chunk- Likestd::for_each, but explicitly splits the input range into chunks then exposes the chunked parallelism. A user-specified chunk constructor is called for each parallel chunk then its output is passed to each loop iteration. Useful for workloads that need an expensive workspace that can be reused between iterations, but not simultaneously by all iterations in parallel.poolstl::pluggable_sort- Likestd::sort, but allows specification of sequential sort method. To parallelize pdqsort:pluggable_sort(par, v.begin(), v.end(), pdqsort).
PoolSTL provides:
poolstl::par: Substitute forstd::execution::par. Parallelized using a thread pool.poolstl::seq: Substitute forstd::execution::seq. Simply calls the regular (non-policy) overload.poolstl::par_if(): Choose parallel or sequential at runtime. See below.
In short, use poolstl::par to make your code parallel. Complete example:
#include <iostream>
#include <poolstl/poolstl.hpp>
int main() {
std::vector<int> v = {0, 1, 2, 3, 4, 5};
auto sum = std::reduce(poolstl::par, vec.cbegin(), vec.cend());
// ^^^^^^^^^^^^
// Add this to make your code parallel.
std::cout << "Sum=" << sum << std::endl;
return 0;
}The thread pool used by poolstl::par is managed internally by poolSTL. It is started on first use.
Use your own thread pool
with poolstl::par.on(pool) for control over thread count, startup/shutdown, etc.:
task_thread_pool::task_thread_pool pool{4}; // 4 threads
std::reduce(poolstl::par.on(pool), vec.begin(), vec.end());Sometimes the choice whether to parallelize or not should be made at runtime. For example, small datasets may not amortize the cost of starting threads, while large datasets do and should be parallelized.
Use poolstl::par_if to select between par and seq at runtime:
bool is_parallel = vec.size() > 10000;
std::reduce(poolstl::par_if(is_parallel), vec.begin(), vec.end());Use poolstl::par_if(is_parallel, pool) to control the thread pool used by par, if selected.
std::vector<int> vec = {0, 1, 2, 3, 4, 5};
// Parallel for-each
std::for_each(poolstl::par, vec.begin(), vec.end(), [](auto& value) {
std::cout << value; // loop body
});using poolstl::iota_iter;
// parallel for loop
std::for_each(poolstl::par, iota_iter<int>(0), iota_iter<int>(100), [](auto i) {
std::cout << i; // loop body
});std::vector<int> vec = {5, 2, 1, 3, 0, 4};
std::sort(poolstl::par, vec.begin(), vec.end());Each release publishes a single-file amalgamated poolstl.hpp. Simply copy this into your project.
Build requirements:
- Clang and GCC 8 or older: require
-lpthreadto use C++11 threads. - Emscripten: compile and link with
-pthreadto use C++11 threads. See docs.
include(FetchContent)
FetchContent_Declare(
poolSTL
GIT_REPOSITORY https://github.com/alugowski/poolSTL
GIT_TAG main
GIT_SHALLOW TRUE
)
FetchContent_MakeAvailable(poolSTL)
target_link_libraries(YOUR_TARGET poolSTL::poolSTL)Alternatively copy or checkout the repo into your project and:
add_subdirectory(poolSTL)See benchmark/ to compare poolSTL against the standard sequential implementation, and (if available) the
native std::execution::par implementation.
Results on an M1 Pro (6 power, 2 efficiency cores), with GCC 13:
-------------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations
-------------------------------------------------------------------------------------------------------
all_of()/real_time 19.9 ms 19.9 ms 35
all_of(poolstl::par)/real_time 3.47 ms 0.119 ms 198
all_of(std::execution::par)/real_time 3.45 ms 3.25 ms 213
find_if()/needle_percentile:5/real_time 0.988 ms 0.987 ms 712
find_if()/needle_percentile:50/real_time 9.87 ms 9.86 ms 71
find_if()/needle_percentile:100/real_time 19.7 ms 19.7 ms 36
find_if(poolstl::par)/needle_percentile:5/real_time 0.405 ms 0.050 ms 1730
find_if(poolstl::par)/needle_percentile:50/real_time 1.85 ms 0.096 ms 393
find_if(poolstl::par)/needle_percentile:100/real_time 3.64 ms 0.102 ms 193
find_if(std::execution::par)/needle_percentile:5/real_time 0.230 ms 0.220 ms 3103
find_if(std::execution::par)/needle_percentile:50/real_time 1.75 ms 1.60 ms 410
find_if(std::execution::par)/needle_percentile:100/real_time 3.51 ms 3.24 ms 204
for_each()/real_time 94.6 ms 94.6 ms 7
for_each(poolstl::par)/real_time 18.7 ms 0.044 ms 36
for_each(std::execution::par)/real_time 15.3 ms 12.9 ms 46
sort()/real_time 603 ms 602 ms 1
sort(poolstl::par)/real_time 112 ms 6.64 ms 6
sort(std::execution::par)/real_time 113 ms 102 ms 6
pluggable_sort(poolstl::par, ..., pdqsort)/real_time 71.7 ms 6.67 ms 10
transform()/real_time 95.0 ms 94.9 ms 7
transform(poolstl::par)/real_time 17.4 ms 0.037 ms 38
transform(std::execution::par)/real_time 15.3 ms 13.2 ms 45
exclusive_scan()/real_time 33.7 ms 33.7 ms 21
exclusive_scan(poolstl::par)/real_time 11.6 ms 0.095 ms 55
exclusive_scan(std::execution::par)/real_time 19.8 ms 15.3 ms 32
reduce()/real_time 15.2 ms 15.2 ms 46
reduce(poolstl::par)/real_time 4.06 ms 0.044 ms 169
reduce(std::execution::par)/real_time 3.38 ms 3.16 ms 214
USE AT YOUR OWN RISK! THIS IS A HACK!
Two-line hack for missing compiler support. A no-op on compilers with support.
If POOLSTL_STD_SUPPLEMENT is defined then poolSTL will check for native compiler support.
If not found then poolSTL will alias its poolstl::par as std::execution::par:
#define POOLSTL_STD_SUPPLEMENT
#include <poolstl/poolstl.hpp>Now just use std::execution::par as normal, and poolSTL will fill in as necessary. See supplement_test.cpp.
Example use case: You can link against TBB, so you'll use native support on GCC 9+, Clang, MSVC, etc. PoolSTL will fill in automatically on GCC <9 and Apple Clang.
Example use case 2: You'd prefer to use the TBB version, but don't want to fail on systems that don't have it.
Simply use the supplement as above, but have your build system (CMake, meson, etc.) check for TBB.
If not found, define POOLSTL_STD_SUPPLEMENT_NO_INCLUDE and the supplement will not #include <execution> (and neither should your code!),
thus dropping the TBB link requirement. The poolSTL supplement fills in.
See the supplement section of tests/CMakeLists.txt for an example.