Add support for waitpkg and reporting count in clock and stats

include/nvsl/clock.hh

@@ -158,8 +158,7 @@ namespace nvsl {
 #endif
 
       size_t ops_per_iter = total_ops / raw_values.size();
-      ss << this->summarize()
-         << "ops: " << total_ops
+      ss << this->summarize() << "ops: " << total_ops
          << "\nops/s: " << (total_ops * (1000000000)) / ((double)this->ns())
          << "\ntime/op: "
          << ns_to_hr_clk((size_t)(this->ns() / (double)total_ops))

include/nvsl/stats.hh

@@ -90,7 +90,7 @@ namespace nvsl {
      */
     void notify_sample() {
       if constexpr (periodic_stat_dump) {
-        if (sample_count++ > STAT_DUMP_PERIOD) {
+        if (sample_count++ % STAT_DUMP_PERIOD == 0) {
           const std::filesystem::path STAT_DUMP_DIR("/tmp/");
           const auto ofstream_flags = std::ios::out | std::ios::trunc;
           std::ofstream dump_file(STAT_DUMP_DIR / this->dump_file_name(),
@@ -210,7 +210,9 @@ namespace nvsl {
      */
     std::string str() const {
       std::stringstream ss;
-      ss << stat_name + ".bucket_count: " << bucket_cnt << "\t# " + stat_desc
+      ss << stat_name + ".sample_count: " << total() << "\t# " + stat_desc
+         << "\n"
+         << stat_name + ".bucket_count: " << bucket_cnt << "\t# " + stat_desc
          << "\n"
          << stat_name + ".bucket_min: " << bucket_min << "\t# " + stat_desc
          << "\n"
@@ -219,6 +221,8 @@ namespace nvsl {
          << stat_name + ".bucket_size: " << bucket_sz << "\t# " + stat_desc
          << "\n"
          << stat_name + ".mean: " << sum / total() << "\t# " + stat_desc << "\n"
+         << stat_name + ".mean_per_k: " << (sum / total()) / 1000
+         << "\t# " + stat_desc << " (= mean/1000)\n"
          << stat_name + ".underflow_count: " << underflow_cnt
          << "\t# " + stat_desc << "\n"
          << stat_name + ".overflow_count: " << overflow_cnt

include/nvsl/waitpkg.hh

@@ -0,0 +1,31 @@
+#include <immintrin.h>
+#include <x86intrin.h>
+
+#pragma once
+
+namespace nvsl {
+  enum class TpauseType {
+    DeepSleep = 0,  // C0.2
+    LightSleep = 1, // C0.1
+  };
+
+  static inline void tpause(size_t wait_cycles, TpauseType type) {
+    const uint64_t current_tsc = _rdtsc();
+    const uint64_t final_tsc = current_tsc + wait_cycles;
+
+    _tpause(static_cast<int>(type), final_tsc);
+  }
+
+  static inline void wait(void *addr, size_t wait_cycles, TpauseType type) {
+    _umonitor(addr);
+
+    while (true) {
+      const uint64_t current_tsc = _rdtsc();
+      const uint64_t final_tsc = current_tsc + wait_cycles;
+
+      const auto expired = _umwait(static_cast<int>(type), final_tsc);
+
+      if (!expired) break;
+    };
+  }
+} // namespace nvsl