Merge pull request #236 from intel/develop

Add option to set min auto-tuner trial time.

src/common/common_utils.cpp

@@ -43,7 +43,7 @@ namespace yask {
     // for numbers above 9 (at least up to 99).
 
     // Format: "major.minor.patch".
-    const string version = "3.01.05";
+    const string version = "3.02.00";
 
     string yask_get_version_string() {
         return version;

src/kernel/lib/auto_tuner.cpp

@@ -150,14 +150,14 @@ namespace yask {
         STATE_VARS(this);
         if (tune_mini_blks())
             DEBUG_MSG(_name << ": best-mini-block-size: " <<
-                      target_sizes().makeDimValStr(" * "));
+                      target_sizes().removeDim(step_posn).makeDimValStr(" * "));
         else
             DEBUG_MSG(_name << ": best-block-size: " <<
-                      target_sizes().makeDimValStr(" * ") << endl <<
+                      target_sizes().removeDim(step_posn).makeDimValStr(" * ") << endl <<
                       _name << ": mini-block-size: " <<
-                      _settings->_mini_block_sizes.makeDimValStr(" * "));
+                      _settings->_mini_block_sizes.removeDim(step_posn).makeDimValStr(" * "));
         DEBUG_MSG(_name << ": sub-block-size: " <<
-                  _settings->_sub_block_sizes.makeDimValStr(" * "));
+                  _settings->_sub_block_sizes.removeDim(step_posn).makeDimValStr(" * "));
     }
 
     // Access settings.
@@ -238,7 +238,7 @@ namespace yask {
         timer.clear();
         idx_t steps = steps_done;
         steps_done = 0;
-        
+
         // Leave if done.
         if (done)
             return;
@@ -247,15 +247,22 @@ namespace yask {
         if (!nullop)
             return;
 
-        // Cumulative stats.
+        // Cumulative stats and rate.
         csteps += steps;
         ctime += etime;
-
+        double rate = (ctime > 0.) ? (double(csteps) / ctime) : 0.;
+        double min_secs = _settings->_tuner_min_secs;
+        TRACE_MSG(_name << " eval() callback: " << steps << " step(s) in " <<
+                  etime << " secs; " << csteps << " step(s) in " <<
+                  ctime << " secs (" << rate <<
+                  " steps/sec) cumulative; best-rate = " << best_rate <<
+                  "; min-secs = " << min_secs);
+
         // Still in warmup?
         if (in_warmup) {
 
             // Warmup not done?
-            if (ctime < warmup_secs && csteps < warmup_steps)
+            if (ctime < max(warmup_secs, min_secs) && csteps < warmup_steps)
                 return;
 
             // Done.
@@ -294,12 +301,7 @@ namespace yask {
             return;
         }
 
-        // Calc perf.
-        double rate = (ctime > 0.) ? (double(csteps) / ctime) : 0.;
-        TRACE_MSG(_name << ": " <<
-                  makeNumStr(rate) << " steps/sec (" <<
-                  csteps << " steps(s) in " << makeNumStr(ctime) <<
-                  " secs)");
+        // Determine whether we've done enough.
         bool rate_ok = false;
 
         // If the current rate is much less than the best,
@@ -315,14 +317,6 @@ namespace yask {
         if (!rate_ok)
             return;
 
-        // Print progress and reset vars for next time.
-        DEBUG_MSG(_name << ": search-dist=" << radius << ": " <<
-                  makeNumStr(rate) << " steps/sec (" <<
-                  csteps << " steps(s) in " << makeNumStr(ctime) <<
-                  " secs) with size " << target_sizes().makeDimValStr(" * "));
-        csteps = 0;
-        ctime = 0.;
-
         // Save result.
         results[target_sizes()] = rate;
         bool is_better = rate > best_rate;
@@ -332,6 +326,16 @@ namespace yask {
             better_neigh_found = true;
         }
 
+        // Print progress and reset vars for next time.
+        DEBUG_MSG(_name << ": search-dist=" << radius << ": " <<
+                  makeNumStr(rate) << " steps/sec (" <<
+                  csteps << " steps(s) in " << makeNumStr(ctime) <<
+                  " secs) with size " <<
+                  target_sizes().removeDim(step_posn).makeDimValStr(" * ") <<
+                  (is_better ? " -- best so far" : ""));
+        csteps = 0;
+        ctime = 0.;
+
         // At this point, we have gathered perf info on the current settings.
         // Now, we need to determine next unevaluated point in search space.
         while (true) {

src/kernel/lib/auto_tuner.hpp

@@ -51,7 +51,6 @@ namespace yask {
         double warmup_steps = 100;
         double warmup_secs = 0.5; // end warmup when either warmup_steps OR warmup_secs is reached.
         idx_t min_steps = 100;
-        double min_secs = 0.25; // eval when either min_steps OR min_secs is reached.
         double cutoff = 0.8;   // can stop eval if current rate < best rate * cutoff;
         idx_t max_radius = 8;   // starting search radius.
         idx_t min_dist = 4;     // min distance to move in any direction per eval.

src/kernel/lib/settings.cpp

@@ -424,6 +424,11 @@ namespace yask {
                            "Adjust block sizes *during* normal operation to tune for performance. "
                            "May cause varying performance between steps.",
                            _do_auto_tune));
+        parser.add_option(new CommandLineParser::DoubleOption
+                          ("auto_tune_min_secs",
+                           "Minimum seconds to run trial during auto-tuning for trial settings to be "
+                           "considered better than the existing best.",
+                           _tuner_min_secs));
         parser.add_option(new CommandLineParser::BoolOption
                           ("auto_tune_mini_blocks",
                            "Apply the auto-tuner to mini-block sizes instead of block sizes. "

src/kernel/lib/settings.hpp

@@ -260,6 +260,7 @@ namespace yask {
         bool _do_auto_tune = false;    // whether to do auto-tuning.
         bool _tune_mini_blks = false; // auto-tune mini-blks instead of blks.
         bool _allow_pack_tuners = false; // allow per-pack tuners when possible.
+        double _tuner_min_secs = 0.25;   // min time to run tuner for new better setting.
 
         // Debug.
         bool force_scalar = false; // Do only scalar ops.

src/kernel/lib/utils.cpp

@@ -428,6 +428,28 @@ namespace yask {
         return false;
     }
 
+    // Get one double value from args[argi].
+    // On failure, print msg using string from args[argi-1] and exit.
+    // On success, increment argi and return value.
+    double CommandLineParser::OptionBase::_double_val(const vector<string>& args,
+                                                     int& argi)
+    {
+        if (size_t(argi) >= args.size() || args[argi].length() == 0) {
+            THROW_YASK_EXCEPTION("Error: no argument for option '" + args[argi - 1] + "'");
+        }
+
+        const char* nptr = args[argi].c_str();
+        char* endptr = 0;
+        double val = strtod(nptr, &endptr);
+        if (val == HUGE_VAL || val == -HUGE_VAL || *endptr != '\0') {
+            THROW_YASK_EXCEPTION("Error: argument for option '" + args[argi - 1] +
+                                 "' is not a valid floating-point number");
+        }
+
+        argi++;
+        return val;
+    }
+
     // Get one idx_t value from args[argi].
     // On failure, print msg using string from args[argi-1] and exit.
     // On success, increment argi and return value.
@@ -472,11 +494,29 @@ namespace yask {
             (_val ? "true" : "false") << "." << endl;
     }
 
+    // Check for a double option.
+    bool CommandLineParser::DoubleOption::check_arg(const std::vector<std::string>& args,
+                                                    int& argi) {
+        if (_check_arg(args, argi, _name)) {
+            _val = _double_val(args, argi);
+            return true;
+        }
+        return false;
+    }
+
+    // Print help on a double option.
+    void CommandLineParser::DoubleOption::print_help(ostream& os,
+                                                     int width) const {
+        _print_help(os, _name + " <floating-point number>", width);
+        os << _help_leader << _current_value_str <<
+            _val << "." << endl;
+    }
+
     // Check for an int option.
     bool CommandLineParser::IntOption::check_arg(const std::vector<std::string>& args,
                                                  int& argi) {
         if (_check_arg(args, argi, _name)) {
-            _val = (int)_idx_val(args, argi); // TODO: check for under/overflow.
+            _val = (int)_idx_val(args, argi); // TODO: check for over/underflow.
             return true;
         }
         return false;

src/kernel/lib/utils.hpp

@@ -345,6 +345,10 @@ namespace yask {
             virtual bool _check_arg(const std::vector<std::string>& args, int& argi,
                                     const std::string& str) const;
 
+            // Get one double value from args[argi++].
+            // Exit on failure.
+            virtual double _double_val(const std::vector<std::string>& args, int& argi);
+
             // Get one idx_t value from args[argi++].
             // Exit on failure.
             virtual idx_t _idx_val(const std::vector<std::string>& args, int& argi);
@@ -407,6 +411,21 @@ namespace yask {
             virtual bool check_arg(const std::vector<std::string>& args, int& argi);
         };
 
+        // An allowed double option.
+        class DoubleOption : public OptionBase {
+            double& _val;
+
+        public:
+            DoubleOption(const std::string& name,
+                      const std::string& help_msg,
+                      double& val) :
+                OptionBase(name, help_msg), _val(val) { }
+
+            virtual void print_help(std::ostream& os,
+                                    int width) const;
+            virtual bool check_arg(const std::vector<std::string>& args, int& argi);
+        };
+
         // An allowed idx_t option.
         class IdxOption : public OptionBase {
             idx_t& _val;

src/kernel/yask_main.cpp

@@ -272,7 +272,7 @@ int main(int argc, char** argv)
         yk_factory kfac;
         yask_output_factory yof;
 
-        // Parse options once just to get vars needed for env.
+        // Parse custom options once just to get vars needed for env.
         MySettings opts1(nullptr);
         opts1.parse(argc, argv);