Add NCUObserver

Author: csbnw

kernel_tuner/core.py

@@ -23,7 +23,7 @@
 from kernel_tuner.backends.opencl import OpenCLFunctions
 from kernel_tuner.backends.hip import HipFunctions
 from kernel_tuner.observers.nvml import NVMLObserver
-from kernel_tuner.observers.observer import ContinuousObserver, OutputObserver
+from kernel_tuner.observers.observer import ContinuousObserver, OutputObserver, PrologueObserver
 
 try:
     import torch
@@ -319,6 +319,7 @@ def __init__(
         self.use_nvml = False
         self.continuous_observers = []
         self.output_observers = []
+        self.prologue_observers = []
         if observers:
             for obs in observers:
                 if isinstance(obs, NVMLObserver):
@@ -328,7 +329,8 @@ def __init__(
                     self.continuous_observers.append(obs.continuous_observer)
                 if isinstance(obs, OutputObserver):
                     self.output_observers.append(obs)
-
+                if isinstance(obs, PrologueObserver):
+                    self.prologue_observers.append(obs)
 
         self.iterations = iterations
 
@@ -346,6 +348,12 @@ def benchmark_default(self, func, gpu_args, threads, grid, result):
             obs for obs in self.dev.observers if not isinstance(obs, ContinuousObserver)
         ]
 
+        for obs in self.prologue_observers:
+            obs.before_start()
+            self.dev.run_kernel(func, gpu_args, threads, grid)
+            self.dev.synchronize()
+            obs.after_finish()
+
         self.dev.synchronize()
         for _ in range(self.iterations):
             for obs in observers:

kernel_tuner/observers/__init__.py

@@ -1 +1 @@
-from .observer import BenchmarkObserver, IterationObserver, ContinuousObserver, OutputObserver
+from .observer import BenchmarkObserver, IterationObserver, ContinuousObserver, OutputObserver, PrologueObserver

kernel_tuner/observers/ncu.py

@@ -0,0 +1,36 @@
+from kernel_tuner.observers import PrologueObserver
+
+try:
+    import nvmetrics
+except (ImportError):
+    nvmetrics = None
+    pass
+
+class NCUObserver(PrologueObserver):
+    """``NCUObserver`` measures performance counters.
+
+    """
+
+    def __init__(self, metrics=None):
+        """Create a new ``NCUObserver``.
+
+        :param metrics: The metrics to observe. This should be a list of strings.
+                        You can use ``ncu --query-metrics`` to get a list of valid metrics.
+        """
+
+        if not nvmetrics:
+            print("NCUObserver is not available.")
+
+        self.metrics = metrics
+        self.results = dict()
+
+    def before_start(self):
+        if nvmetrics:
+            nvmetrics.measureMetricsStart(self.metrics)
+
+    def after_finish(self):
+        if nvmetrics:
+            self.results = nvmetrics.measureMetricsStop()
+
+    def get_results(self):
+        return dict(zip(self.metrics, self.results))

kernel_tuner/observers/observer.py

@@ -57,4 +57,15 @@ def process_output(self, answer, output):
         """
         pass
 
+class PrologueObserver(BenchmarkObserver):
+    """Observer that measures something in a seperate kernel invocation prior to the normal benchmark."""
 
+    @abstractmethod
+    def before_start(self):
+        """before start is called before the kernel starts"""
+        pass
+
+    @abstractmethod
+    def after_finish(self):
+        """after finish is called after the kernel has finished execution"""
+        pass