ncnn刚发布不久,博主在ios下尝试编译。
遇上了openmp的编译问题。
寻找各种解决方案无果,亲自操刀。
采用std::thread 替换 openmp。
ncnn项目地址:
https://github.com/Tencent/ncnn
后来询问ncnn的作者才知道在ios下的编译方法。
至此,当时的临时方案 采用std::thread 替换 openmp。
想想也许在一些特定情况下还是比较适用的,当前方便两者之间进行切换验证。
抽空写了一个示例项目。
项目地址:
https://github.com/cpuimage/ParallelFor
贴上完整代码:
#include <stdio.h> #include <stdlib.h> #include <iostream> #if defined(_OPENMP) // compile with: /openmp #include <omp.h> auto const epoch = omp_get_wtime(); double now() { return omp_get_wtime() - epoch; }; #else #include <chrono> auto const epoch = std::chrono::steady_clock::now(); double now() { return std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - epoch).count() / 1000.0; }; #endif template<typename FN> double bench(const FN &fn) { auto took = -now(); return (fn(), took + now()); } #include <functional> #if defined(_OPENMP) # include <omp.h> #else #include <thread> #include <vector> #endif #ifdef _OPENMP static int processorCount = static_cast<int>(omp_get_num_procs()); #else static int processorCount = static_cast<int>(std::thread::hardware_concurrency()); #endif static void ParallelFor(int inclusiveFrom, int exclusiveTo, std::function<void(size_t)> func) { #if defined(_OPENMP) #pragma omp parallel for num_threads(processorCount) for (int i = inclusiveFrom; i < exclusiveTo; ++i) { func(i); } return; #else if (inclusiveFrom >= exclusiveTo) return; static size_t thread_cnt = 0; if (thread_cnt == 0) { thread_cnt = std::thread::hardware_concurrency(); } size_t entry_per_thread = (exclusiveTo - inclusiveFrom) / thread_cnt; if (entry_per_thread < 1) { for (int i = inclusiveFrom; i < exclusiveTo; ++i) { func(i); } return; } std::vector<std::thread> threads; int start_idx, end_idx; for (start_idx = inclusiveFrom; start_idx < exclusiveTo; start_idx += entry_per_thread) { end_idx = start_idx + entry_per_thread; if (end_idx > exclusiveTo) end_idx = exclusiveTo; threads.emplace_back([&](size_t from, size_t to) { for (size_t entry_idx = from; entry_idx < to; ++entry_idx) func(entry_idx); }, start_idx, end_idx); } for (auto& t : threads) { t.join(); } #endif } void test_scale(int i, double* a, double* b) { a[i] = 4 * b[i]; } int main() { int N = 10000; double* a2 = (double*)calloc(N, sizeof(double)); double* a1 = (double*)calloc(N, sizeof(double)); double* b = (double*)calloc(N, sizeof(double)); if (a1 == NULL || a2 == NULL || b == NULL) { if (a1) { free(a1); }if (a2) { free(a2); }if (b) { free(b); } return -1; } for (int i = 0; i < N; i++) { a1[i] = i; a2[i] = i; b[i] = i; } double beforeTime = bench([&] { for (int i = 0; i < N; i++) { test_scale(i, a1, b); } }); std::cout << " before: " << int(beforeTime * 1000) << "ms" << std::endl; double afterTime = bench([&] { ParallelFor(0, N, [a2, b](size_t i) { test_scale(i, a2, b); }); }); std::cout << " after: " << int(afterTime * 1000) << "ms" << std::endl; for (int i = 0; i < N; i++) { if (a1[i] != a2[i]) { printf("error %f : %f ", a1[i], a2[i]); getchar(); } } free(a1); free(a2); free(b); getchar(); return 0; }
要使用OPENMP,加个编译选项/openmp 或者定义一下 _OPENMP 即可。
建议c++11编译。
示例代码比较简单。
ncnn代码修改例子如下:
#pragma omp parallel for for (int q=0; q<channels; q++) { const Mat m = src.channel(q); Mat borderm = dst.channel(q); copy_make_border_image(m, borderm, top, left, type, v); }
改为
ParallelFor(0, channels, [&](int q) { { const Mat m = src.channel(q); Mat borderm = dst.channel(q); copy_make_border_image(m, borderm, top, left, type, v); }});
本来计划抽点时间把ncnn整体都改一下,发个修改版本出来。
想想还是把做法贴出来,给有需求的人吧。
自己动手丰衣足食。
若有其他相关问题或者需求也可以邮件联系俺探讨。
邮箱地址是:
gaozhihan@vip.qq.com