use tbb::concurrent_unordered_map instead of std::map so data races d…

…o not happen across threads in profiling

stan/math/rev/core/profiling.hpp

@@ -7,6 +7,7 @@
 #include <stan/math/rev/meta.hpp>
 #include <stan/math/rev/fun/value_of.hpp>
 #include <stan/math/prim/err.hpp>
+#include <tbb/concurrent_unordered_map.h>
 #include <iostream>
 #include <sstream>
 #include <thread>
@@ -115,7 +116,24 @@ class profile_info {
 
 using profile_key = std::pair<std::string, std::thread::id>;
 
-using profile_map = std::map<profile_key, profile_info>;
+
+namespace internal {
+  struct hash_profile_key {
+    std::size_t operator()(const profile_key& key) const {
+      return std::hash<std::string>()(key.first)
+             ^ std::hash<std::thread::id>()(key.second);
+    }
+  };
+  struct equal_profile_key {
+    bool operator()(const profile_key& lhs, const profile_key& rhs) const {
+      return lhs.first == rhs.first && lhs.second == rhs.second;
+    }
+  };
+
+}
+
+using profile_map = tbb::concurrent_unordered_map<profile_key,
+  profile_info, internal::hash_profile_key, internal::equal_profile_key>;
 
 /**
  * Profiles C++ lines where the object is in scope.

test/unit/math/rev/core/profiling_test.cpp

@@ -6,31 +6,31 @@
 TEST(Profiling, double_basic) {
   using stan::math::profile;
   using stan::math::var;
-  stan::math::profile_map profiles;
+  stan::math::profile_map prof_map;
   double a = 3.0, b = 2.0, c;
   {
-    profile<double> p1("p1", profiles);
+    profile<double> p1("p1", prof_map);
     c = log(exp(a)) * log(exp(b));
     std::chrono::milliseconds timespan(10);
     std::this_thread::sleep_for(timespan);
   }
   {
-    profile<int> p1("p1", profiles);
+    profile<int> p1("p1", prof_map);
     c = log(exp(a)) * log(exp(b));
     std::chrono::milliseconds timespan(10);
     std::this_thread::sleep_for(timespan);
   }
 
   stan::math::profile_key key = {"p1", std::this_thread::get_id()};
   EXPECT_NEAR(c, 6.0, 1E-8);
-  EXPECT_EQ(profiles[key].get_chain_stack_used(), 0);
-  EXPECT_EQ(profiles[key].get_nochain_stack_used(), 0);
-  EXPECT_FLOAT_EQ(profiles[key].get_rev_time(), 0.0);
-  EXPECT_EQ(profiles[key].get_num_rev_passes(), 0);
-  EXPECT_EQ(profiles[key].get_num_fwd_passes(), 2);
-  EXPECT_EQ(profiles[key].get_num_no_AD_fwd_passes(), 2);
-  EXPECT_EQ(profiles[key].get_num_AD_fwd_passes(), 0);
-  EXPECT_TRUE(profiles[key].get_fwd_time() > 0.0);
+  EXPECT_EQ(prof_map[key].get_chain_stack_used(), 0);
+  EXPECT_EQ(prof_map[key].get_nochain_stack_used(), 0);
+  EXPECT_FLOAT_EQ(prof_map[key].get_rev_time(), 0.0);
+  EXPECT_EQ(prof_map[key].get_num_rev_passes(), 0);
+  EXPECT_EQ(prof_map[key].get_num_fwd_passes(), 2);
+  EXPECT_EQ(prof_map[key].get_num_no_AD_fwd_passes(), 2);
+  EXPECT_EQ(prof_map[key].get_num_AD_fwd_passes(), 0);
+  EXPECT_TRUE(prof_map[key].get_fwd_time() > 0.0);
 }
 
 TEST(Profiling, var_basic) {