Time spent setting NVML parameters (clock & memory frequency, power) goes to framework time instead of benchmark time

Author: fjwillemsen

kernel_tuner/core.py

@@ -391,12 +391,8 @@ def benchmark_continuous(self, func, gpu_args, threads, grid, result, duration):
         for obs in self.continuous_observers:
             result.update(obs.get_results())
 
-    def benchmark(self, func, gpu_args, instance, verbose, objective):
-        """benchmark the kernel instance"""
-        logging.debug("benchmark " + instance.name)
-        logging.debug("thread block dimensions x,y,z=%d,%d,%d", *instance.threads)
-        logging.debug("grid dimensions x,y,z=%d,%d,%d", *instance.grid)
-
+    def set_nvml_parameters(self, instance):
+        """Set the NVML parameters. Avoids setting time leaking into benchmark time."""
         if self.use_nvml:
             if "nvml_pwr_limit" in instance.params:
                 new_limit = int(
@@ -409,6 +405,15 @@ def benchmark(self, func, gpu_args, instance, verbose, objective):
             if "nvml_mem_clock" in instance.params:
                 self.nvml.mem_clock = instance.params["nvml_mem_clock"]
 
+    def benchmark(self, func, gpu_args, instance, verbose, objective, skip_nvml_setting=False):
+        """Benchmark the kernel instance."""
+        logging.debug("benchmark " + instance.name)
+        logging.debug("thread block dimensions x,y,z=%d,%d,%d", *instance.threads)
+        logging.debug("grid dimensions x,y,z=%d,%d,%d", *instance.grid)
+
+        if self.use_nvml and not skip_nvml_setting:
+            self.set_nvml_parameters(instance)
+
         # Call the observers to register the configuration to be benchmarked
         for obs in self.dev.observers:
             obs.register_configuration(instance.params)
@@ -577,11 +582,15 @@ def compile_and_benchmark(self, kernel_source, gpu_args, params, kernel_options,
 
                 # benchmark
                 if func:
+                    # setting the NVML parameters here avoids this time from leaking into the benchmark time, ends up in framework time instead
+                    if self.use_nvml:
+                        self.set_nvml_parameters(instance)
                     start_benchmark = time.perf_counter()
                     result.update(
-                        self.benchmark(func, gpu_args, instance, verbose, to.objective)
+                        self.benchmark(func, gpu_args, instance, verbose, to.objective, skip_nvml_setting=False)
                     )
                     last_benchmark_time = 1000 * (time.perf_counter() - start_benchmark)
+                    print(f"Benchmark time: {last_benchmark_time}")
 
             except Exception as e:
                 # dump kernel sources to temp file

kernel_tuner/runners/sequential.py

@@ -100,7 +100,7 @@ def run(self, parameter_space, tuning_options):
                 params = process_metrics(params, tuning_options.metrics)
 
             # get the framework time by estimating based on other times
-            total_time = 1000 * (perf_counter() - self.start_time) - warmup_time
+            total_time = 1000 * ((perf_counter() - self.start_time) - warmup_time)    # TODO is it valid that we deduct the warmup time here?
             params['strategy_time'] = self.last_strategy_time
             params['framework_time'] = max(total_time - (params['compile_time'] + params['verification_time'] + params['benchmark_time'] + params['strategy_time']), 0)
             params['timestamp'] = str(datetime.now(timezone.utc))