| 113 | |
| 114 | template <class Pool, class RunThread> |
| 115 | void my_main(int argc, char** argv, exec::numa_policy policy = exec::get_numa_policy()) |
| 116 | { |
| 117 | int nthreads = static_cast<int>(std::thread::hardware_concurrency()); |
| 118 | if (argc > 1) |
| 119 | { |
| 120 | nthreads = std::atoi(argv[1]); |
| 121 | } |
| 122 | std::size_t total_scheds = 10'000'000; |
| 123 | #ifndef STDEXEC_NO_MONOTONIC_BUFFER_RESOURCE |
| 124 | std::vector<std::unique_ptr<char, numa_deleter>> buffers; |
| 125 | #endif |
| 126 | std::optional<Pool> pool{}; |
| 127 | if constexpr (std::same_as<Pool, exec::static_thread_pool>) |
| 128 | { |
| 129 | pool.emplace(nthreads, exec::bwos_params{}, policy); |
| 130 | } |
| 131 | else |
| 132 | { |
| 133 | pool.emplace(nthreads); |
| 134 | } |
| 135 | std::barrier<> barrier(nthreads + 1); |
| 136 | std::vector<std::thread> threads; |
| 137 | std::atomic<bool> stop{false}; |
| 138 | #ifndef STDEXEC_NO_MONOTONIC_BUFFER_RESOURCE |
| 139 | std::size_t buffer_size = 2000 << 20; |
| 140 | for (std::size_t i = 0; std::cmp_less(i, nthreads); ++i) |
| 141 | { |
| 142 | exec::numa_allocator<char> alloc(policy.thread_index_to_node(i)); |
| 143 | buffers.push_back(std::unique_ptr<char, numa_deleter>{ |
| 144 | alloc.allocate(buffer_size), |
| 145 | numa_deleter{.size_ = buffer_size, .allocator_ = alloc} |
| 146 | }); |
| 147 | } |
| 148 | #endif |
| 149 | for (std::size_t i = 0; std::cmp_less(i, nthreads); ++i) |
| 150 | { |
| 151 | threads.emplace_back(RunThread{}, |
| 152 | std::ref(*pool), |
| 153 | total_scheds, |
| 154 | i, |
| 155 | std::ref(barrier), |
| 156 | #ifndef STDEXEC_NO_MONOTONIC_BUFFER_RESOURCE |
| 157 | std::span<char>{buffers[i].get(), buffer_size}, |
| 158 | #endif |
| 159 | std::ref(stop), |
| 160 | policy); |
| 161 | } |
| 162 | std::size_t nRuns = 100; |
| 163 | std::size_t warmup = 1; |
| 164 | std::vector<std::chrono::steady_clock::time_point> starts(nRuns); |
| 165 | std::vector<std::chrono::steady_clock::time_point> ends(nRuns); |
| 166 | for (std::size_t i = 0; i < nRuns; ++i) |
| 167 | { |
| 168 | barrier.arrive_and_wait(); |
| 169 | starts[i] = std::chrono::steady_clock::now(); |
| 170 | barrier.arrive_and_wait(); |
| 171 | ends[i] = std::chrono::steady_clock::now(); |
| 172 | if (i < warmup) |
nothing calls this directly
no test coverage detected