Merge branch 'master' into directives

Author: isazi

examples/cuda/vector_add_tegra_observer.py

@@ -0,0 +1,46 @@
+#!/usr/bin/env python
+"""This is the minimal example from the README"""
+
+import json
+
+import numpy
+from kernel_tuner import tune_kernel
+from kernel_tuner.observers.tegra import TegraObserver
+
+def tune():
+
+    kernel_string = """
+    __global__ void vector_add(float *c, float *a, float *b, int n) {
+        int i = blockIdx.x * block_size_x + threadIdx.x;
+        if (i<n) {
+            c[i] = a[i] + b[i];
+        }
+    }
+    """
+
+    size = 800000
+
+    a = numpy.random.randn(size).astype(numpy.float32)
+    b = numpy.random.randn(size).astype(numpy.float32)
+    c = numpy.zeros_like(b)
+    n = numpy.int32(size)
+
+    args = [c, a, b, n]
+
+    tune_params = dict()
+    tune_params["block_size_x"] = [128+64*i for i in range(15)]
+
+    tegraobserver = TegraObserver(["core_freq"])
+
+    metrics = dict()
+    metrics["f"] = lambda p: p["core_freq"]
+
+    results, env = tune_kernel("vector_add", kernel_string, size, args, tune_params, observers=[tegraobserver], metrics=metrics)
+
+    print(results)
+
+    return results
+
+
+if __name__ == "__main__":
+    tune()

kernel_tuner/backends/hip.py

@@ -19,6 +19,7 @@
     "bool": ctypes.c_bool,
     "int8": ctypes.c_int8,
     "int16": ctypes.c_int16,
+    "float16": ctypes.c_int16,
     "int32": ctypes.c_int32,
     "int64": ctypes.c_int64,
     "uint8": ctypes.c_uint8,

kernel_tuner/core.py

@@ -20,9 +20,8 @@
 from kernel_tuner.backends.nvcuda import CudaFunctions
 from kernel_tuner.backends.opencl import OpenCLFunctions
 from kernel_tuner.backends.compiler import CompilerFunctions
-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.tegra import TegraObserver
 from kernel_tuner.observers.observer import ContinuousObserver, OutputObserver, PrologueObserver
 
 try:
@@ -307,8 +306,9 @@ def __init__(
             )
         self.dev = dev
 
-        # look for NVMLObserver in observers, if present, enable special tunable parameters through nvml
+        # look for NVMLObserver and TegraObserver in observers, if present, enable special tunable parameters through nvml/tegra
         self.use_nvml = False
+        self.use_tegra = False
         self.continuous_observers = []
         self.output_observers = []
         self.prologue_observers = []
@@ -317,6 +317,9 @@ def __init__(
                 if isinstance(obs, NVMLObserver):
                     self.nvml = obs.nvml
                     self.use_nvml = True
+                if isinstance(obs, TegraObserver):
+                    self.tegra = obs.tegra
+                    self.use_tegra = True
                 if hasattr(obs, "continuous_observer"):
                     self.continuous_observers.append(obs.continuous_observer)
                 if isinstance(obs, OutputObserver):
@@ -373,6 +376,7 @@ def benchmark_default(self, func, gpu_args, threads, grid, result):
         for obs in self.benchmark_observers:
             result.update(obs.get_results())
 
+
     def benchmark_continuous(self, func, gpu_args, threads, grid, result, duration):
         """Benchmark continuously for at least 'duration' seconds"""
         iterations = int(np.ceil(duration / (result["time"] / 1000)))
@@ -396,6 +400,7 @@ def benchmark_continuous(self, func, gpu_args, threads, grid, result, duration):
         for obs in self.continuous_observers:
             result.update(obs.get_results())
 
+
     def set_nvml_parameters(self, instance):
         """Set the NVML parameters. Avoids setting time leaking into benchmark time."""
         if self.use_nvml:
@@ -410,6 +415,11 @@ def set_nvml_parameters(self, instance):
             if "nvml_mem_clock" in instance.params:
                 self.nvml.mem_clock = instance.params["nvml_mem_clock"]
 
+        if self.use_tegra:
+            if "tegra_gr_clock" in instance.params:
+                self.tegra.gr_clock = instance.params["tegra_gr_clock"]
+
+
     def benchmark(self, func, gpu_args, instance, verbose, objective, skip_nvml_setting=False):
         """Benchmark the kernel instance."""
         logging.debug("benchmark " + instance.name)

kernel_tuner/observers/nvml.py

@@ -323,7 +323,7 @@ def __init__(
         save_all=False,
         nvidia_smi_fallback=None,
         use_locked_clocks=False,
-        continous_duration=1,
+        continuous_duration=1,
     ):
         """Create an NVMLObserver."""
         if nvidia_smi_fallback:
@@ -355,7 +355,7 @@ def __init__(
         if any([obs in self.needs_power for obs in observables]):
             self.measure_power = True
             power_observables = [obs for obs in observables if obs in self.needs_power]
-            self.continuous_observer = NVMLPowerObserver(power_observables, self, self.nvml, continous_duration)
+            self.continuous_observer = ContinuousObserver("nvml", power_observables, self, continuous_duration=continuous_duration)
 
         # remove power observables
         self.observables = [obs for obs in observables if obs not in self.needs_power]
@@ -373,6 +373,10 @@ def __init__(
         self.during_obs = [obs for obs in observables if obs in ["core_freq", "mem_freq", "temperature"]]
         self.iteration = {obs: [] for obs in self.during_obs}
 
+    def read_power(self):
+        """ Return power in Watt """
+        return self.nvml.pwr_usage() / 1e3
+
     def before_start(self):
         # clear results of the observables for next measurement
         self.iteration = {obs: [] for obs in self.during_obs}
@@ -428,75 +432,6 @@ def get_results(self):
         return averaged_results
 
 
-class NVMLPowerObserver(ContinuousObserver):
-    """Observer that measures power using NVML and continuous benchmarking."""
-
-    def __init__(self, observables, parent, nvml_instance, continous_duration=1):
-        self.parent = parent
-        self.nvml = nvml_instance
-
-        supported = ["power_readings", "nvml_power", "nvml_energy"]
-        for obs in observables:
-            if obs not in supported:
-                raise ValueError(f"Observable {obs} not in supported: {supported}")
-        self.observables = observables
-
-        # duration in seconds
-        self.continuous_duration = continous_duration
-
-        self.power = 0
-        self.energy = 0
-        self.power_readings = []
-        self.t0 = 0
-
-        # results from the last iteration-based benchmark
-        self.results = None
-
-    def before_start(self):
-        self.parent.before_start()
-        self.power = 0
-        self.energy = 0
-        self.power_readings = []
-
-    def after_start(self):
-        self.parent.after_start()
-        self.t0 = time.perf_counter()
-
-    def during(self):
-        self.parent.during()
-        power_usage = self.nvml.pwr_usage()
-        timestamp = time.perf_counter() - self.t0
-        # only store the result if we get a new measurement from NVML
-        if len(self.power_readings) == 0 or (
-            self.power_readings[-1][1] != power_usage or timestamp - self.power_readings[-1][0] > 0.01
-        ):
-            self.power_readings.append([timestamp, power_usage])
-
-    def after_finish(self):
-        self.parent.after_finish()
-        # safeguard in case we have no measurements, perhaps the kernel was too short to measure anything
-        if not self.power_readings:
-            return
-
-        # convert to seconds from milliseconds
-        execution_time = self.results["time"] / 1e3
-        self.power = np.median([d[1] / 1e3 for d in self.power_readings])
-        self.energy = self.power * execution_time
-
-    def get_results(self):
-        results = self.parent.get_results()
-        keys = list(results.keys())
-        for key in keys:
-            results["pwr_" + key] = results.pop(key)
-        if "nvml_energy" in self.observables:
-            results["nvml_energy"] = self.energy
-        if "nvml_power" in self.observables:
-            results["nvml_power"] = self.power
-        if "power_readings" in self.observables:
-            results["power_readings"] = self.power_readings
-        return results
-
-
 # High-level Helper functions
 
 

kernel_tuner/observers/observer.py

@@ -1,5 +1,6 @@
 from abc import ABC, abstractmethod
-
+import time
+import numpy as np
 
 class BenchmarkObserver(ABC):
     """Base class for Benchmark Observers"""
@@ -44,8 +45,78 @@ class IterationObserver(BenchmarkObserver):
 
 
 class ContinuousObserver(BenchmarkObserver):
-    pass
+    """Generic observer that measures power while and continuous benchmarking.
+
+        To support continuous benchmarking an Observer should support:
+        a .read_power() method, which the ContinuousObserver can call to read power in Watt
+    """
+    def __init__(self, name, observables, parent, continuous_duration=1):
+        self.parent = parent
+        self.name = name
+
+        supported = [self.name + "_power", self.name + "_energy", "power_readings"]
+        for obs in observables:
+            if obs not in supported:
+                raise ValueError(f"Observable {obs} not in supported: {supported}")
+        self.observables = observables
+
+        # duration in seconds
+        self.continuous_duration = continuous_duration
+
+        self.power = 0
+        self.energy = 0
+        self.power_readings = []
+        self.t0 = 0
+
+        # results from the last iteration-based benchmark
+        # these are set by the benchmarking function of Kernel Tuner before
+        # the continuous observer is called.
+        self.results = None
+
+    def before_start(self):
+        self.parent.before_start()
+        self.power = 0
+        self.energy = 0
+        self.power_readings = []
+
+    def after_start(self):
+        self.parent.after_start()
+        self.t0 = time.perf_counter()
 
+    def during(self):
+        self.parent.during()
+        power_usage = self.parent.read_power()
+        timestamp = time.perf_counter() - self.t0
+        # only store the result if we get a new measurement from the GPU
+        if len(self.power_readings) == 0 or (
+            self.power_readings[-1][1] != power_usage
+            or timestamp - self.power_readings[-1][0] > 0.01
+        ):
+            self.power_readings.append([timestamp, power_usage])
+
+    def after_finish(self):
+        self.parent.after_finish()
+        # safeguard in case we have no measurements, perhaps the kernel was too short to measure anything
+        if not self.power_readings:
+            return
+
+        # convert to seconds from milliseconds
+        execution_time = self.results["time"] / 1e3
+        self.power = np.median([d[1] for d in self.power_readings])
+        self.energy = self.power * execution_time
+
+    def get_results(self):
+        results = self.parent.get_results()
+        keys = list(results.keys())
+        for key in keys:
+            results["pwr_" + key] = results.pop(key)
+        if self.name + "_power" in self.observables:
+            results[self.name + "_power"] = self.power
+        if self.name + "_energy" in self.observables:
+            results[self.name + "_energy"] = self.energy
+        if "power_readings" in self.observables:
+            results["power_readings"] = self.power_readings
+        return results
 
 class OutputObserver(BenchmarkObserver):
     """Observer that can verify or measure something about the output produced by a kernel."""

kernel_tuner/observers/pmt.py

@@ -1,6 +1,6 @@
 import numpy as np
 
-from kernel_tuner.observers.observer import BenchmarkObserver
+from kernel_tuner.observers.observer import BenchmarkObserver, ContinuousObserver
 
 # check if pmt is installed
 try:
@@ -28,9 +28,25 @@ class PMTObserver(BenchmarkObserver):
 
     :type observables: string,list/dictionary
 
+
+    :param use_continuous_observer:
+        Boolean to control whether or not to measure power/energy using
+        Kernel Tuner's continuous benchmarking mode. This improves measurement
+        accuracy when using internal power sensors, such as NVML or ROCM,
+        which have limited sampling frequency and might return averages
+        instead of instantaneous power readings. Default value: False.
+
+    :type use_continuous_observer: boolean
+
+
+    :param continuous_duration:
+        Number of seconds to measure continuously for.
+
+    :type continuous_duration: scalar
+
     """
 
-    def __init__(self, observable=None):
+    def __init__(self, observable=None, use_continuous_observer=False, continuous_duration=1):
         if not pmt:
             raise ImportError("could not import pmt")
 
@@ -54,6 +70,9 @@ def __init__(self, observable=None):
         self.begin_states = [None] * len(self.pms)
         self.initialize_results(self.pm_names)
 
+        if use_continuous_observer:
+            self.continuous_observer = PMTContinuousObserver("pmt", [], self, continuous_duration=continuous_duration)
+
     def initialize_results(self, pm_names):
         self.results = dict()
         for pm_name in pm_names:
@@ -82,3 +101,39 @@ def get_results(self):
         averages = {key: np.average(values) for key, values in self.results.items()}
         self.initialize_results(self.pm_names)
         return averages
+
+
+class PMTContinuousObserver(ContinuousObserver):
+    """Generic observer that measures power while and continuous benchmarking.
+
+        To support continuous benchmarking an Observer should support:
+        a .read_power() method, which the ContinuousObserver can call to read power in Watt
+    """
+    def before_start(self):
+        """ Override default method in ContinuousObserver """
+        pass
+
+    def after_start(self):
+        self.parent.after_start()
+
+    def during(self):
+        """ Override default method in ContinuousObserver """
+        pass
+
+    def after_finish(self):
+        self.parent.after_finish()
+
+    def get_results(self):
+        average_kernel_execution_time_ms = self.results["time"]
+
+        averages = {key: np.average(values) for key, values in self.results.items()}
+        self.parent.initialize_results(self.parent.pm_names)
+
+        # correct energy measurement, because current _energy number is collected over the entire duration
+        # we estimate energy as the average power over the continuous duration times the kernel execution time
+        for pm_name in self.parent.pm_names:
+            energy_result_name = f"{pm_name}_energy"
+            power_result_name = f"{pm_name}_power"
+            averages[energy_result_name] = averages[power_result_name] * (average_kernel_execution_time_ms / 1e3)
+
+        return averages