modified average finding to correct for repeats and changed 1000ms time limit to config parameter

include/ck/host_utility/kernel_launch.hpp

@@ -45,9 +45,10 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
             float total_time = 0;
 
             hip_check_error(hipEventElapsedTime(&total_time, start, stop));
-            total_time /= 10;
+            total_time /= stream_config.nrepeat_;
             stream_config.cold_niters_ =
-                (1000.0 / total_time); // we need longer runtime to ramp up the clk on MI300s
+                (stream_config.time_limit_ms /
+                 total_time); // we need longer runtime to ramp up the clk on MI300s
             stream_config.nrepeat_ = stream_config.cold_niters_;
         }
 #endif
@@ -148,9 +149,10 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
             float total_time = 0;
 
             hip_check_error(hipEventElapsedTime(&total_time, start, stop));
-            total_time /= 10;
+            total_time /= stream_config.nrepeat_;
             stream_config.cold_niters_ =
-                (1000.0 / total_time); // we need longer runtime to ramp up the clk on MI300s
+                (stream_config.nrepeat_ /
+                 total_time); // we need longer runtime to ramp up the clk on MI300s
             stream_config.nrepeat_ = stream_config.cold_niters_;
         }
 #endif

include/ck/stream_config.hpp

@@ -13,4 +13,5 @@ struct StreamConfig
     int log_level_            = 0;
     mutable int cold_niters_  = 5;
     mutable int nrepeat_      = 50;
+    mutable int time_limit_ms = 1000; // for timing MI300 runs
 };