Add a benchmark for parallel loops

Author: matt-42

benchmarks/parallel_for.cc

@@ -0,0 +1,112 @@
+#include <iostream>
+#include <vpp/boxNd.hh>
+#include <vpp/vpp.hh>
+
+#include "get_time.hh"
+
+using namespace vpp;
+
+template <typename I, typename F>
+void fast_pixel_wise(I img, F f)
+{
+  int nr = img.nrows();
+#pragma omp parallel for
+  for (int r = 0; r < nr; r++)
+  {
+    int* cur = &img(vint2(r, 0));
+    // vint2 cur_(r,0);
+    int* end = cur + img.nrows();
+    int c = 0;
+    while (cur != end)
+    {
+      f(cur, r, c);
+      //f(cur, cur_[0], cur_[1]);
+      ++cur;
+      // cur_[1]++;
+      c++;
+    }
+  }
+
+}
+
+int main()
+{
+
+  image2d<int> img(2000,2000);
+
+  int K = 400;
+  double time;
+
+
+  // check
+  vpp::pixel_wise(img) << [] (auto&& p)
+  {
+    p = 42;
+  };
+
+  for (auto p : img.domain())
+  {
+    if (img(p) != 42)
+      std::cout << "error at " << p.transpose() << " -> " << img(p) << std::endl;
+  }
+
+  // Cache warm up.
+  for (int k = 0; k < K; k++)
+    fast_pixel_wise(img, [] (int* p, int r, int c)
+    {
+      *p = r + c;
+    });
+
+  time = get_time_in_seconds();
+  for (int k = 0; k < K; k++)
+  {
+    int nr = img.nrows();
+#pragma omp parallel for
+    for (int r = 0; r < nr; r++)
+    {
+      int* cur = &img(vint2(r, 0));
+      int* end = cur + img.nrows();
+      int c = 0;
+      while (cur != end)
+      {
+        *cur = r + c;
+        ++cur;
+        c++;
+      }
+
+    }
+  }
+  double ref_time = get_time_in_seconds() - time;
+
+
+  time = get_time_in_seconds();
+  for (int k = 0; k < K; k++)
+    fast_pixel_wise(img, [] (int* p, int r, int c)
+    {
+      *p = r + c;
+    });
+  double ref2_time = get_time_in_seconds() - time;
+
+  time = get_time_in_seconds();
+  for (int k = 0; k < K; k++)
+    for (auto& p : img)
+      p = p.coord()[0] + p.coord()[1];
+  double ii_time = get_time_in_seconds() - time;
+
+  time = get_time_in_seconds();
+  for (int k = 0; k < K; k++)
+    vpp::pixel_wise(img, img.domain()) << [] (auto& p, auto& p2)
+    {
+      p = p2[0] + p2[1];
+    };
+  double pw_time = get_time_in_seconds() - time;
+
+  std::cout << "image iterator: " << std::endl;
+  std::cout << "ref_time (ms): " << 1000. * ref_time / K << std::endl;
+  std::cout << "ref2_time (ms): " << 1000. * ref2_time / K << std::endl;
+  std::cout << "ii_time (ms): " << 1000. * ii_time / K << std::endl;
+  std::cout << "image iterator overhead: " << 100. * ii_time / ref_time - 100. << "%" << std::endl;
+  std::cout << "pw_time (ms): " << 1000. * pw_time / K << std::endl;
+  //std::cout << "domain iteration overhead: " << 100. * id_time / ref_time - 100. << "%" << std::endl;
+
+}