Better throttle recovery delay

Author: gevtushenko

docs/cli_help.md

@@ -140,13 +140,13 @@
 
 * `--throttle-threshold <value>`
   * Set the GPU throttle threshold as percentage of the peak clock rate.
-  * Default is 0.75 (75%).
+  * Default is 75%.
   * Applies to the most recent `--benchmark`, or all benchmarks if specified
     before any `--benchmark` arguments.
 
 * `--throttle-recovery-delay <value>`
   * Set the GPU throttle recovery delay in seconds.
-  * Default is 0.5 seconds.
+  * Default is 0.05 seconds.
   * Applies to the most recent `--benchmark`, or all benchmarks if specified
     before any `--benchmark` arguments.
 

nvbench/benchmark_base.cuh

@@ -309,8 +309,8 @@ protected:
   nvbench::float64_t m_skip_time{-1.};
   nvbench::float64_t m_timeout{15.};
 
-  nvbench::float32_t m_throttle_threshold{0.75f};     // [% of peak SM clock rate]
-  nvbench::float32_t m_throttle_recovery_delay{0.0f}; // [seconds]
+  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;

nvbench/detail/measure_cold.cuh

@@ -18,24 +18,24 @@
 
 #pragma once
 
+#include <cuda_runtime.h>
+
 #include <nvbench/blocking_kernel.cuh>
 #include <nvbench/cpu_timer.cuh>
 #include <nvbench/cuda_call.cuh>
 #include <nvbench/cuda_timer.cuh>
+#include <nvbench/detail/gpu_frequency.cuh>
+#include <nvbench/detail/kernel_launcher_timer_wrapper.cuh>
+#include <nvbench/detail/l2flush.cuh>
+#include <nvbench/detail/statistics.cuh>
 #include <nvbench/device_info.cuh>
 #include <nvbench/exec_tag.cuh>
 #include <nvbench/launch.cuh>
 #include <nvbench/stopping_criterion.cuh>
 
-#include <nvbench/detail/kernel_launcher_timer_wrapper.cuh>
-#include <nvbench/detail/l2flush.cuh>
-#include <nvbench/detail/statistics.cuh>
-#include <nvbench/detail/gpu_frequency.cuh>
-
-#include <cuda_runtime.h>
-
 #include <utility>
 #include <vector>
+
 #include "nvbench/types.cuh"
 
 namespace nvbench
@@ -91,7 +91,7 @@ protected:
   nvbench::blocking_kernel m_blocker;
 
   nvbench::criterion_params m_criterion_params;
-  nvbench::stopping_criterion_base& m_stopping_criterion;
+  nvbench::stopping_criterion_base &m_stopping_criterion;
   nvbench::detail::gpu_frequency m_gpu_frequency;
 
   bool m_disable_blocking_kernel{false};
@@ -102,8 +102,8 @@ protected:
   nvbench::float64_t m_skip_time{};
   nvbench::float64_t m_timeout{};
 
-  nvbench::float32_t m_throttle_threshold{0.75f};     // [% of peak SM clock rate]
-  nvbench::float32_t m_throttle_recovery_delay{0.0f}; // [seconds]
+  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

@@ -1071,7 +1071,7 @@ try
   }
   else if (prop_arg == "--throttle-threshold")
   {
-    bench.set_throttle_threshold(static_cast<nvbench::float32_t>(value));
+    bench.set_throttle_threshold(static_cast<nvbench::float32_t>(value) / 100.0f);
   }
   else if (prop_arg == "--throttle-recovery-delay")
   {

nvbench/state.cuh

@@ -338,8 +338,8 @@ private:
   nvbench::float64_t m_skip_time;
   nvbench::float64_t m_timeout;
 
-  nvbench::float32_t m_throttle_threshold{0.75f};     // [% of peak SM clock rate]
-  nvbench::float32_t m_throttle_recovery_delay{0.0f}; // [seconds]
+  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.