eval/BFP9000: record csv

This also refactors a bit when we take timestamps.

eval/BFP9000.cpp

@@ -51,11 +51,11 @@ static std::uint64_t number_of_pulses;
 static std::uint64_t pulse_target_sleep_ns;
 
 struct PulseTrace {
-	chrn::time_point<Clock> sleep_finish;
 	std::uint64_t work_ns;
 	chrn::time_point<Clock> work_start;
 	chrn::time_point<Clock> work_finish;
 	std::uint64_t sleep_ns;
+	chrn::time_point<Clock> sleep_start;
 };
 
 class Pulser {
@@ -67,6 +67,7 @@ class Pulser {
 
  public:
 	std::vector<PulseTrace> traces;
+	chrn::time_point<Clock> pulse_start;
 
  private:
 	static void busyWait(std::uint64_t work_ns) {
@@ -119,9 +120,9 @@ class Pulser {
 
 		if (trace) {
 			traces = std::vector<PulseTrace>(number_of_pulses);
-			traces[0].sleep_finish = Clock::now();
 		}
 
+		pulse_start = Clock::now();
 		for (std::uint64_t i = 0; i < number_of_pulses; ++i) {
 			const auto work_ns = getNanos(work_dist, remaining_work_ns, pulse_target_work_ns, i);
 			if (trace) {
@@ -135,12 +136,10 @@ class Pulser {
 			const auto sleep_ns = getNanos(sleep_dist, remaining_sleep_ns, pulse_target_sleep_ns, i);
 			if (trace) {
 				traces[i].sleep_ns = sleep_ns;
-				auto next_i = i + 1;
-				if (next_i < number_of_pulses) {
-					traces[next_i].sleep_finish = Clock::now() + chrn::nanoseconds(sleep_ns);
-				}
+				traces[i].sleep_start = Clock::now();
 			}
 			emper::nanosleep(sleep_ns);
+			// TODO: record actual sleep finish?
 			remaining_sleep_ns -= sleep_ns;
 		}
 	}
@@ -204,6 +203,7 @@ static void bfp9000() {
 	std::optional<std::ofstream> latencies_file_strm;
 	std::optional<std::ofstream> work_file_strm;
 	std::optional<std::ofstream> sleep_file_strm;
+	std::optional<std::ofstream> csv_file_strm;
 	if (result_dir) {
 		auto latencies_file = *result_dir / "latencies";
 		latencies_file_strm = std::ofstream(latencies_file);
@@ -213,27 +213,48 @@ static void bfp9000() {
 
 		auto sleep_file = *result_dir / "sleep";
 		sleep_file_strm = std::ofstream(sleep_file);
+
+		auto csv_file = *result_dir / "csv";
+		csv_file_strm = std::ofstream(csv_file);
 	}
 
 	for (auto& pulser : pulsers) {
 		std::uint64_t pulser_work_ns = 0;
 		std::uint64_t pulser_sleep_ns = 0;
+
+		auto previous_sleep_finish = pulser.pulse_start;
 		for (auto& trace : pulser.traces) {
 			pulser_work_ns += trace.work_ns;
 			pulser_sleep_ns += trace.sleep_ns;
 
-			auto pulse_latency =
-					chrn::duration_cast<chrn::nanoseconds>(trace.work_start - trace.sleep_finish);
+			auto sleep_finish = previous_sleep_finish;
+			previous_sleep_finish = trace.work_finish + chrn::nanoseconds(trace.sleep_ns);
+
+			auto pulse_latency = chrn::duration_cast<chrn::nanoseconds>(trace.work_start - sleep_finish);
 
 			assert(pulse_latency.count());
 			latencies.push_back(pulse_latency);
 
 			if (!result_dir) continue;
 
+			auto work_wall_clock =
+					chrn::duration_cast<chrn::nanoseconds>(trace.work_finish - trace.work_start);
+			auto work_speedup =
+					static_cast<double>(trace.work_ns) / static_cast<double>(work_wall_clock.count());
+
 			// NOLINTBEGIN(bugprone-unchecked-optional-access)
 			*latencies_file_strm << pulse_latency.count() << std::endl;
 			*work_file_strm << trace.work_ns << std::endl;
 			*sleep_file_strm << trace.sleep_ns << std::endl;
+			// clang-format off
+			*csv_file_strm << sleep_finish
+										 << ", " << trace.work_start
+										 << ", " << pulse_latency
+										 << ", " << trace.work_finish
+										 << ", " << work_wall_clock
+										 << ", " << work_speedup
+										;
+			// clang-format on
 			// NOLINTEND(bugprone-unchecked-optional-access)
 		}