diff --git a/llvm/projects/gpu_profiler/include/profiler.h b/llvm/projects/gpu_profiler/include/profiler.h
index 9468554833fa82fd3b3c8c2ab85a393e21e92c80..33f36dba7d7281de84dd4a155762559f24e5c52b 100644
--- a/llvm/projects/gpu_profiler/include/profiler.h
+++ b/llvm/projects/gpu_profiler/include/profiler.h
@@ -8,7 +8,7 @@
#include <thread>
#include <vector>
-// Reads power rails at runtime and computes the GPU and DDR energy within a window
+// Reads power rails at runtime and computes the specified device and DDR energy within a window
// of time, which is delimitered by the calls to resume_profiler() and pause_profiler()
//
// IMPORTANT: Must call pause_profiler() to kill the profiler thread
@@ -28,6 +28,16 @@ public:
~Profiler();
+ // FIXME: Probably change the enum name to something else.
+ // Expose the devices available
+ enum Device {
+ CPU = 0,
+ GPU = 1 << 0,
+ SOC = 1 << 1,
+ SYS = 1 << 2,
+ NONE = 1 << 3,
+ };
+
// Reinitializes boolean vars used for control flow and launches the profiler
// thread. DOES NOT reset other internal data structures.
void start_profiler();
@@ -39,11 +49,11 @@ public:
// Stops profiler by putting profiler thread to sleep
void pause_profiler();
- // Gets the delta time and total GPU and DDR energy between the last two
+ // Gets the delta time and total CPU/GPU and DDR energy between the last two
// calls to resume_profiler and pause_profiler
//
// Returns this as a pair of <delta time in milliseconds, energy>
- std::pair<double, double> get_time_energy() const;
+ std::pair<double, double> get_time_energy(Device dev = GPU) const;
// Resets all internal data structures, including the vector storing all power_readings.
void reset();
@@ -76,11 +86,12 @@ private:
// An individual power reading
struct PowerReading {
std::chrono::time_point<clock_type> time_;
- double cpu_;
- double gpu_;
+ double dev_;
double ddr_;
- double soc_;
- double sys_;
+ Device dev_type_;
+
+ // Initialize to avoid any undefined behavior
+ //PowerReading() : dev_type_(NONE), dev_(0), ddr_(0) {}
};
// Stores all power readings and is cleared only when reset() is called
@@ -111,13 +122,14 @@ private:
std::thread profiler_thread_;
- // Obtain's a single power reading from the GPU and DDR rails
- void obtain_power_reading();
-
+ // Obtain's a single power reading from the device and DDR rails
+ void obtain_power_reading(Device dev);
+
// Pins the given thread to the specified core
void pin_thread(std::thread &t, const unsigned core) const;
// Runs the profiler thread, keeping it alive by wrapping the functionality
// in an infinite loop
void run_profiler();
+
};
diff --git a/llvm/projects/gpu_profiler/src/profiler.cpp b/llvm/projects/gpu_profiler/src/profiler.cpp
index cae8823fa23056d44973f3b23d16c4c2f28e89f7..de84067d946451acce032ae3cdaaef667c588a4a 100644
--- a/llvm/projects/gpu_profiler/src/profiler.cpp
+++ b/llvm/projects/gpu_profiler/src/profiler.cpp
@@ -43,39 +43,38 @@ void Profiler::start_profiler(){
// Resumes the profiling of whatever executable's currently running
// DOES NOT reset any data
void Profiler::resume_profiler() {
- {
- std::unique_lock<std::mutex> mutex_lock(mutex_);
- if (should_run_profiler_){
- std::cout << "WARNING: resume_profiler was already called\n";
- }
- //std::cout<<"RESUME RESUME RESUME RESUME\n";
- should_run_profiler_ = true;
- start_time_ = clock_type::now();
- }
+ std::unique_lock<std::mutex> mutex_lock(mutex_);
+ if (should_run_profiler_){
+ std::cout << "WARNING: resume_profiler was already called\n";
+ }
+ //std::cout<<"RESUME RESUME RESUME RESUME\n";
+ should_run_profiler_ = true;
+ start_time_ = clock_type::now();
cond_var_.notify_one();
}
// Stops profiler by putting profiler thread to sleep
void Profiler::pause_profiler() {
- {
- std::unique_lock<std::mutex> mutex_lock(mutex_);
- if (!should_run_profiler_){
- std::cout << "WARNING: pause_profiler was already called\n";
- }
- //std::cout<<"PAUSE PAUSE PAUSE PAUSE\n";
- should_run_profiler_ = false;
- stop_time_ = clock_type::now();
+ std::unique_lock<std::mutex> mutex_lock(mutex_);
+ if (!should_run_profiler_){
+ std::cout << "WARNING: pause_profiler was already called\n";
}
+ //std::cout<<"PAUSE PAUSE PAUSE PAUSE\n";
+ should_run_profiler_ = false;
+ stop_time_ = clock_type::now();
cond_var_.notify_one();
}
-// Gets the delta time and total GPU and DDR energy between the last two
+// Gets the delta time and total CPU/GPU and DDR energy between the last two
// calls to resume_profiler and pause_profiler
//
// Returns this as a pair of <delta time in milliseconds, energy>
-std::pair<double, double> Profiler::get_time_energy() const {
- std::unique_lock<std::mutex> mutex_lock(vector_mutex_); // MUST use a mutex
+std::pair<double, double> Profiler::get_time_energy(Device dev) const {
+ // We support taking CPU/GPU readings only
+ if(dev != CPU && dev != GPU)
+ return std::make_pair(-1, -1);
+ std::unique_lock<std::mutex> mutex_lock(vector_mutex_); // MUST use a mutex
double total_energy = 0.0;
if (on_jetson_) {
//std::cout<<"power readings size"<<power_readings_.size()<<'\n';
@@ -83,7 +82,8 @@ std::pair<double, double> Profiler::get_time_energy() const {
for (size_t i = 0; i < power_readings_.size(); i++){
const auto& reading = power_readings_[i];
std::chrono::duration<double> duration_secs = reading.time_ - prev_time;
- total_energy += (reading.gpu_ + reading.ddr_);// * duration_secs.count();
+ if(reading.dev_type_ == dev)
+ total_energy += (reading.dev_ + reading.ddr_);// * duration_secs.count();
prev_time = reading.time_;
}
}
@@ -110,10 +110,13 @@ void Profiler::stop_profiler() {
profiler_thread_.join();
}
-// Obtain's a single power reading from the GPU and DDR rails
-void Profiler::obtain_power_reading() {
- std::unique_lock<std::mutex> mutex_lock(vector_mutex_); // MUST use a mutex
+// Obtain's a single power reading from the CPU/GPU and DDR rails
+void Profiler::obtain_power_reading(Device dev) {
+ // We support taking CPU/GPU readings only
+ if(dev != CPU && dev != GPU)
+ return;
+ std::unique_lock<std::mutex> mutex_lock(vector_mutex_); // MUST use a mutex
PowerReading reading;
// The order matters here. All the reads have to happen together first
@@ -122,16 +125,22 @@ void Profiler::obtain_power_reading() {
// the different rails.
reading.time_ = clock_type::now();
if (on_jetson_){
- gpu_stream_ >> reading.gpu_;
- ddr_stream_ >> reading.ddr_;
-
- gpu_stream_.seekg(0);
- ddr_stream_.seekg(0);
-
+ // FIXME: Use switch-case in the future.
+ if(dev == CPU) {
+ cpu_stream_ >> reading.dev_;
+ ddr_stream_ >> reading.ddr_;
+ cpu_stream_.seekg(0);
+ } else {
+ gpu_stream_ >> reading.dev_;
+ ddr_stream_ >> reading.ddr_;
+ gpu_stream_.seekg(0);
+ }
+ ddr_stream_.seekg(0);
} else {
- reading.gpu_ = 0.0;
+ reading.dev_ = 0.0;
reading.ddr_ = 0.0;
}
+ reading.dev_type_ = dev;
power_readings_.push_back(reading);
}
@@ -165,7 +174,8 @@ void Profiler::run_profiler(){
if (should_stop_profiler_) {
break;
}
- obtain_power_reading();
+ obtain_power_reading(GPU);
+ obtain_power_reading(CPU);
}
}