Squashed commit of the following:

commit 2ac5b61d544fb46a7f74e62729882ba15100bf1f
Author: Carl Pearson <pearson@illinois.edu>
Date:   Fri Sep 20 14:53:10 2019 -0500

    changelog in readme

commit 1072efa135c840753892d350985129d6ebc88e7e
Author: Carl Pearson <pearson@illinois.edu>
Date:   Fri Sep 20 14:50:09 2019 -0500

    clean up gpu_monitor example

commit a8b111a84e61d466d1eb3f4ed9133b428f775714
Author: Carl Pearson <pearson@illinois.edu>
Date:   Fri Sep 20 14:47:52 2019 -0500

    add examples/gpu_monitor

commit b0a46680c666c06119d74d462acb2e9a69bf7d85
Author: Carl Pearson <pearson@illinois.edu>
Date:   Fri Sep 20 14:46:12 2019 -0500

    update readme

commit b8d29e4d89b5ca0e7fffe71dc969e44cbbd53ab9
Author: Carl Pearson <pearson@illinois.edu>
Date:   Fri Sep 20 14:41:27 2019 -0500

    add GPU monitor

include/perfect/gpu_monitor.hpp

@@ -0,0 +1,147 @@
+#pragma once
+
+#ifdef __NVCC__
+#ifndef PERFECT_HAS_CUDA
+#define PERFECT_HAS_CUDA
+#endif
+#endif
+
+#ifdef PERFECT_HAS_CUDA
+#include <nvml.h>
+#endif
+
+#include <atomic>
+#include <chrono>
+#include <string>
+#include <thread>
+
+#include <iostream>
+
+#include "perfect/init.hpp"
+
+namespace perfect {
+
+class Monitor {
+public:
+  struct Config {
+    std::atomic<bool> stop;
+    std::atomic<bool> pause;
+    double samplePeriodMs;
+
+    std::atomic<bool> power;
+    std::atomic<bool> utilization;
+    std::atomic<bool> temperature;
+    std::atomic<bool> pstate;
+    std::ostream *stream_;
+
+    Config(std::ostream *stream)
+        : stop(true), pause(false), power(true), samplePeriodMs(100),
+          utilization(true), temperature(true), pstate(true), stream_(stream) {}
+  };
+
+  std::thread worker;
+  Config config;
+
+  Monitor(std::ostream *stream) : config(stream) {}
+
+  static void worker_func(const Config &cfg) {
+
+    nvmlReturn_t ret;
+    nvmlDevice_t device;
+    unsigned int deviceCount;
+
+    nvmlUtilization_t utilization;
+    unsigned int milliwatts;
+    unsigned int temperature;
+    nvmlPstates_t pState;
+
+    ret = nvmlDeviceGetCount(&deviceCount);
+
+    std::chrono::time_point<std::chrono::system_clock> start;
+
+    while (!cfg.stop.load()) {
+      if (std::chrono::time_point<std::chrono::system_clock>() == start) {
+        start = std::chrono::system_clock::now();
+      }
+      if (!cfg.pause.load()) {
+
+        const double elapsed =
+            (std::chrono::system_clock::now() - start).count() / 1e9 * 1e3;
+
+        for (unsigned int i = 0; i < deviceCount; ++i) {
+
+          (*cfg.stream_) << elapsed << "," << i;
+
+          ret = nvmlDeviceGetHandleByIndex(i, &device);
+
+          if (cfg.power.load()) {
+            ret = nvmlDeviceGetPowerUsage(device, &milliwatts);
+            if (ret == NVML_SUCCESS) {
+              (*cfg.stream_) << "," << milliwatts;
+            } else {
+              (*cfg.stream_) << "," << -1;
+            }
+          } else {
+            (*cfg.stream_) << ","
+                           << "x";
+          }
+          if (cfg.utilization.load()) {
+            // period is between 1 second and 1/6 second depending on product
+            ret = nvmlDeviceGetUtilizationRates(device, &utilization);
+            if (ret == NVML_SUCCESS) {
+              (*cfg.stream_)
+                  << "," << utilization.gpu << "," << utilization.memory;
+            } else {
+              (*cfg.stream_) << "," << -1 << "," << -1;
+            }
+          } else {
+            (*cfg.stream_) << ","
+                           << "x"
+                           << ","
+                           << "x";
+          }
+          if (cfg.temperature.load()) {
+            ret = nvmlDeviceGetTemperature(device, NVML_TEMPERATURE_GPU,
+                                           &temperature);
+            if (ret == NVML_SUCCESS) {
+              (*cfg.stream_) << "," << temperature;
+            } else {
+              (*cfg.stream_) << "," << -1;
+            }
+          } else {
+            (*cfg.stream_) << ","
+                           << "x";
+          }
+          if (cfg.pstate.load()) {
+            ret = nvmlDeviceGetPerformanceState(device, &pState);
+            if (ret == NVML_SUCCESS) {
+              (*cfg.stream_) << "," << pState;
+            } else {
+              (*cfg.stream_) << "," << -1;
+            }
+          } else {
+            (*cfg.stream_) << ","
+                           << "x";
+          }
+
+          (*cfg.stream_) << "\n";
+        }
+      }
+      std::this_thread::sleep_for(std::chrono::milliseconds(100));
+    }
+  }
+
+  void start() {
+    config.stop = false;
+    worker = std::thread(worker_func, std::ref(config));
+  }
+  void stop() {
+    config.stop = true;
+    worker.join();
+  }
+
+  void resume() { config.pause.store(false); }
+  void pause() { config.pause.store(true); }
+};
+
+} // namespace perfect

include/perfect/init.hpp

@@ -1,13 +1,17 @@
 #pragma once
 
 #ifdef __NVCC__
+#ifndef PERFECT_HAS_CUDA
 #define PERFECT_HAS_CUDA
 #endif
+#endif
 
 #ifdef PERFECT_HAS_CUDA
 #include <nvml.h>
 #endif
 
+#include "perfect/result.hpp"
+
 namespace perfect {
 
 /*! initialize the benchmark

include/perfect/result.hpp

@@ -3,8 +3,10 @@
 #include <cassert>
 
 #ifdef __NVCC__
+#ifndef PERFECT_HAS_CUDA
 #define PERFECT_HAS_CUDA
 #endif
+#endif
 
 #ifdef PERFECT_HAS_CUDA
 #include <nvml.h>