added option to perform arbitrary number of runs

benchmarks/speed3d.h

@@ -215,9 +215,16 @@ void benchmark_fft(std::array<int,3> size_fft, std::deque<std::string> const &ar
 
     // Print results
     if(me==0){
-        t_max = t_max / (2.0 * ntest);
-        double const fftsize  = static_cast<double>(world.count());
-        double const floprate = 5.0 * batch_size * fftsize * std::log(fftsize) * 1e-9 / std::log(2.0) / t_max;
+        double const fftsize = static_cast<double>(world.count());
+
+        double floprate = 0.0;
+        if (ntest > 0) { // something was tested
+            t_max /= (2.0 * ntest); // time per test, 2 transforms forward/backward
+            floprate = 5.0 * batch_size * fftsize * std::log(fftsize) * 1e-9 / std::log(2.0) / t_max;
+        } else {
+            t_max = 0.0; // nothing was tested
+        }
+
         long long mem_usage = static_cast<long long>(fft.size_inbox()) + static_cast<long long>(fft.size_outbox())
                             + static_cast<long long>(fft.size_workspace());
         mem_usage *= sizeof(output_type);
@@ -232,6 +239,7 @@ void benchmark_fft(std::array<int,3> size_fft, std::deque<std::string> const &ar
         for(int i=0; i<5; i++)
             print_proc_grid(i);
         cout << "\n";
+        cout << "Num runs:     " << ntest << '\n';
         cout << "Time per run: " << t_max << " (s)\n";
         cout << "Performance:  " << floprate << " GFlops/s\n";
         cout << "Memory usage: " << mem_usage << "MB/rank\n";
@@ -351,6 +359,7 @@ int main(int argc, char *argv[]){
      -r2c_dir dir: specifies the r2c direction for the r2c tests, dir must be 0 1 or 2
      -mps: for the cufft backend and multiple gpus, associate the mpi ranks with different cuda devices
      -nX: number of times to repeat the run, accepted variants are -n5 (default), -n1, -n10, -n50
+     -nrunsXYZ: same as -n but allows for a custom number, XYZ must be a non-negative integer, e.g., -nruns17
 
 )help2";
 

test/test_common.h

@@ -356,7 +356,7 @@ bool has_option(std::deque<std::string> const &args, std::string const &opt){
     return false;
 }
 //! \brief Takes the three arguments after \b opt and converts them to an array of ints, throws runtime_error if no arguments or cannot convert.
-std::array<int, 3> get_grid(std::deque<std::string> const &args, std::string const &opt){
+inline std::array<int, 3> get_grid(std::deque<std::string> const &args, std::string const &opt){
     auto iopt = args.begin();
     while(iopt != args.end()){
         if (*iopt == opt){ // found the argument, take the next three entries
@@ -374,7 +374,7 @@ std::array<int, 3> get_grid(std::deque<std::string> const &args, std::string con
     throw std::runtime_error(opt + " not found");
 }
 
-int get_int_arg(std::string const &name, std::deque<std::string> const &args, int default_value = -1){
+inline int get_int_arg(std::string const &name, std::deque<std::string> const &args, int default_value = -1){
     auto iopt = args.begin();
     while(iopt != args.end()){
         if (*iopt == name){
@@ -388,8 +388,27 @@ int get_int_arg(std::string const &name, std::deque<std::string> const &args, in
     }
     return default_value;
 }
-
-int nruns(std::deque<std::string> const &args){
+//! returns the number of runs selected in the args
+inline int nruns(std::deque<std::string> const &args){
+    for(auto &s : args) {
+        std::string::size_type nr = s.find("-nruns");
+        if (nr != 0)
+            continue;
+        // found a string with -nruns, get the number
+        int num = 0;
+        try {
+            num = std::stoi(s.substr(6));
+        } catch(std::invalid_argument &) {
+            std::cerr << "cannot convert '" << s.substr(6) << "' to 'int'\n";
+            throw;
+        } catch(std::out_of_range &) {
+            std::cerr << "provided integer '" << s.substr(6) << "' is too large\n";
+            throw;
+        }
+        if (num < 0)
+            throw std::runtime_error("the number of of runs has to be non-negative");
+        return num;
+    }
     for(auto &s : args)
         if (s == "-n1")
             return 1;