Remove discard on throttle option

Author: gevtushenko

docs/cli_help.md

@@ -133,11 +133,6 @@
   * Applies to the most recent `--benchmark`, or all benchmarks if specified
     before any `--benchmark` arguments.
 
-* `--discard-on-throttle`
-  * Discard measurements if the GPU is throttled.
-  * Applies to the most recent `--benchmark`, or all benchmarks if specified
-    before any `--benchmark` arguments.
-
 * `--throttle-threshold <value>`
   * Set the GPU throttle threshold as percentage of the peak clock rate.
   * Default is 75%.

nvbench/benchmark_base.cuh

@@ -264,13 +264,6 @@ struct benchmark_base
     m_throttle_recovery_delay = throttle_recovery_delay;
   }
 
-  [[nodiscard]] bool get_discard_on_throttle() const { return m_discard_on_throttle; }
-
-  void set_discard_on_throttle(bool discard_on_throttle)
-  {
-    m_discard_on_throttle = discard_on_throttle;
-  }
-
   [[nodiscard]] nvbench::criterion_params &get_criterion_params() { return m_criterion_params; }
   [[nodiscard]] const nvbench::criterion_params &get_criterion_params() const
   {
@@ -311,7 +304,6 @@ protected:
 
   nvbench::float32_t m_throttle_threshold{0.75f};      // [% of peak SM clock rate]
   nvbench::float32_t m_throttle_recovery_delay{0.05f}; // [seconds]
-  bool m_discard_on_throttle{false};
 
   nvbench::criterion_params m_criterion_params;
   std::string m_stopping_criterion{"stdrel"};

nvbench/benchmark_base.cxx

@@ -47,7 +47,6 @@ std::unique_ptr<benchmark_base> benchmark_base::clone() const
   result->m_criterion_params        = m_criterion_params;
   result->m_throttle_threshold      = m_throttle_threshold;
   result->m_throttle_recovery_delay = m_throttle_recovery_delay;
-  result->m_discard_on_throttle     = m_discard_on_throttle;
 
   result->m_stopping_criterion = m_stopping_criterion;
 

nvbench/detail/measure_cold.cu

@@ -26,8 +26,8 @@
 #include <nvbench/summary.cuh>
 
 #include <algorithm>
-#include <limits>
 #include <chrono>
+#include <limits>
 #include <thread>
 
 #include <fmt/format.h>
@@ -48,7 +48,6 @@ measure_cold_base::measure_cold_base(state &exec_state)
     , m_timeout{exec_state.get_timeout()}
     , m_throttle_threshold(exec_state.get_throttle_threshold())
     , m_throttle_recovery_delay(exec_state.get_throttle_recovery_delay())
-    , m_discard_on_throttle(exec_state.get_discard_on_throttle())
 {
   if (m_min_samples > 0)
   {
@@ -96,7 +95,6 @@ void measure_cold_base::record_measurements()
   if (!m_run_once)
   {
     auto peak_clock_rate = static_cast<float>(m_state.get_device()->get_sm_default_clock_rate());
-    m_sm_clock_rates.push_back(peak_clock_rate);
 
     if (m_gpu_frequency.has_throttled(peak_clock_rate, m_throttle_threshold))
     {
@@ -106,14 +104,13 @@ void measure_cold_base::record_measurements()
         auto &printer           = printer_opt_ref.value().get();
         printer.log(nvbench::log_level::warn,
                     fmt::format("GPU throttled below threshold ({:0.2f} MHz / {:0.2f} MHz) "
-                                "({:0.0f}% < {:0.0f}%) on sample {}. {} previous sample and "
-                                "pausing for {}s.",
+                                "({:0.0f}% < {:0.0f}%) on sample {}. Discarding previous sample "
+                                "and pausing for {}s.",
                                 current_clock_rate / 1000000.0f,
                                 peak_clock_rate / 1000000.0f,
                                 100.0f * (current_clock_rate / peak_clock_rate),
                                 100.0f * m_throttle_threshold,
                                 m_total_samples,
-                                m_discard_on_throttle ? "Discarding" : "Keeping",
                                 m_throttle_recovery_delay));
       }
 
@@ -122,11 +119,11 @@ void measure_cold_base::record_measurements()
         std::this_thread::sleep_for(std::chrono::duration<float>(m_throttle_recovery_delay));
       }
 
-      if (m_discard_on_throttle)
-      { // ignore this measurement
-        return;
-      }
+      // ignore this measurement
+      return;
     }
+
+    m_sm_clock_rates.push_back(peak_clock_rate);
   }
 
   // Update and record timers and counters:
@@ -348,8 +345,9 @@ void measure_cold_base::generate_summaries()
     summ.set_string("hint", "frequency");
     summ.set_string("description", "Mean SM clock rate");
     summ.set_string("hide", "Hidden by default.");
-    summ.set_float64("value", nvbench::detail::statistics::compute_mean(m_sm_clock_rates.cbegin(), 
-                                                                        m_sm_clock_rates.cend()));
+    summ.set_float64("value",
+                     nvbench::detail::statistics::compute_mean(m_sm_clock_rates.cbegin(),
+                                                               m_sm_clock_rates.cend()));
   }
 
   // Log if a printer exists:

nvbench/detail/measure_cold.cuh

@@ -104,7 +104,6 @@ protected:
 
   nvbench::float32_t m_throttle_threshold;      // [% of peak SM clock rate]
   nvbench::float32_t m_throttle_recovery_delay; // [seconds]
-  bool m_discard_on_throttle{false};
 
   nvbench::int64_t m_total_samples{};
 

nvbench/option_parser.cu

@@ -431,11 +431,6 @@ void option_parser::parse_range(option_parser::arg_iterator_t first,
       this->lock_gpu_clocks(first[1]);
       first += 2;
     }
-    else if (arg == "--discard-on-throttle")
-    {
-      this->enable_discard_on_throttle();
-      first += 1;
-    }
     else if (arg == "--run-once")
     {
       this->enable_run_once();
@@ -731,18 +726,6 @@ void option_parser::enable_run_once()
   bench.set_run_once(true);
 }
 
-void option_parser::enable_discard_on_throttle()
-{
-  if (m_benchmarks.empty())
-  {
-    m_global_benchmark_args.push_back("--discard-on-throttle");
-    return;
-  }
-
-  benchmark_base &bench = *m_benchmarks.back();
-  bench.set_discard_on_throttle(true);
-}
-
 void option_parser::set_stopping_criterion(const std::string &criterion)
 {
   // If no active benchmark, save args as global.

nvbench/option_parser.cuh

@@ -89,7 +89,6 @@ private:
 
   void set_stopping_criterion(const std::string &criterion);
   void enable_run_once();
-  void enable_discard_on_throttle();
   void disable_blocking_kernel();
 
   void add_benchmark(const std::string &name);

nvbench/state.cuh

@@ -220,13 +220,6 @@ struct state
     m_throttle_recovery_delay = throttle_recovery_delay;
   }
 
-  [[nodiscard]] bool get_discard_on_throttle() const { return m_discard_on_throttle; }
-
-  void set_discard_on_throttle(bool discard_on_throttle)
-  {
-    m_discard_on_throttle = discard_on_throttle;
-  }
-
   /// If a `KernelLauncher` syncs and `nvbench::exec_tag::sync` is not passed
   /// to `state.exec(...)`, a deadlock may occur. If a `blocking_kernel` blocks
   /// for more than `blocking_kernel_timeout` seconds, an error will be printed
@@ -340,7 +333,6 @@ private:
 
   nvbench::float32_t m_throttle_threshold;      // [% of peak SM clock rate]
   nvbench::float32_t m_throttle_recovery_delay; // [seconds]
-  bool m_discard_on_throttle{false};
 
   // Deadlock protection. See blocking_kernel's class doc for details.
   nvbench::float64_t m_blocking_kernel_timeout{30.0};

nvbench/state.cxx

@@ -15,10 +15,9 @@
  *  See the License for the specific language governing permissions and
  *  limitations under the License.
  */
-#include <nvbench/state.cuh>
-
 #include <nvbench/benchmark_base.cuh>
 #include <nvbench/detail/throw.cuh>
+#include <nvbench/state.cuh>
 #include <nvbench/types.cuh>
 
 #include <algorithm>
@@ -43,7 +42,6 @@ state::state(const benchmark_base &bench)
     , m_timeout{bench.get_timeout()}
     , m_throttle_threshold{bench.get_throttle_threshold()}
     , m_throttle_recovery_delay{bench.get_throttle_recovery_delay()}
-    , m_discard_on_throttle{bench.get_discard_on_throttle()}
 {}
 
 state::state(const benchmark_base &bench,
@@ -64,7 +62,6 @@ state::state(const benchmark_base &bench,
     , m_timeout{bench.get_timeout()}
     , m_throttle_threshold{bench.get_throttle_threshold()}
     , m_throttle_recovery_delay{bench.get_throttle_recovery_delay()}
-    , m_discard_on_throttle{bench.get_discard_on_throttle()}
 {}
 
 nvbench::int64_t state::get_int64(const std::string &axis_name) const