Add timing accumulation capability into GpuRegionTimer

Added a TODO to deprecate NB timing structures in favor of new
functionality.

Change-Id: Idb78e5a36a7f372f01378a580a05b928bd728c57

diff --git a/src/gromacs/gpu_utils/gpuregiontimer.h b/src/gromacs/gpu_utils/gpuregiontimer.h
index 0207f96143a614de88b6956d3c6b9a4147657c76..2cbbe575d2bfba23420fa10cf765464c6d4eab51 100644 (file)
--- a/src/gromacs/gpu_utils/gpuregiontimer.h
+++ b/src/gromacs/gpu_utils/gpuregiontimer.h
@@ -104,7 +104,7 @@ template <GpuFramework framework> class GpuRegionTimerImpl
 
 /*! \libinternal \brief
  * This is a GPU region timing wrapper class.
- * It allows for host-side tracking of the execution timespans in GPU code
+ * It allows for host-side tracking of the accumulated execution timespans in GPU code
  * (measuring kernel or transfers duration).
  * It also partially tracks the correctness of the timer state transitions,
  * as far as current implementation allows (see TODO in getLastRangeTime() for a disabled check).
@@ -123,6 +123,10 @@ template <GpuFramework framework> class GpuRegionTimerWrapper
         Stopped
     } debugState_;
 
+    //! The number of times the timespan has been measured
+    unsigned int                  callCount_;
+    //! The accumulated duration of the timespans measured (milliseconds)
+    double                        totalMilliseconds_;
     //! The underlying region timer implementation
     GpuRegionTimerImpl<framework> impl_;
 
@@ -160,10 +164,12 @@ template <GpuFramework framework> class GpuRegionTimerWrapper
                 GMX_ASSERT(debugState_ == TimerState::Recording, error.c_str());
                 debugState_ = TimerState::Stopped;
             }
+            callCount_++;
             impl_.closeTimingRegion(s);
         }
         /*! \brief
-         * Returns the last timespan, and resets the internal timer state.
+         * Accumulates the last timespan of all the events used into the the total duration,
+         * and resets the internal timer state.
          * To be called after closeTimingRegion() and the command stream of the event having been synchronized.
          * \returns The last timespan (in milliseconds).
          */
@@ -186,6 +192,7 @@ template <GpuFramework framework> class GpuRegionTimerWrapper
                 debugState_ = TimerState::Idle;
             }
             double milliseconds = impl_.getLastRangeTime();
+            totalMilliseconds_ += milliseconds;
             return milliseconds;
         }
         /*! \brief Resets the implementation and total time/call count to zeroes. */
@@ -195,8 +202,20 @@ template <GpuFramework framework> class GpuRegionTimerWrapper
             {
                 debugState_ = TimerState::Idle;
             }
+            totalMilliseconds_ = 0.0;
+            callCount_         = 0;
             impl_.reset();
         }
+        /*! \brief Gets total time recorded (in milliseconds). */
+        double getTotalTime() const
+        {
+            return totalMilliseconds_;
+        }
+        /*! \brief Gets total call count recorded. */
+        unsigned int getCallCount() const
+        {
+            return callCount_;
+        }
         /*! \brief
          * Gets a pointer to a new timing event for passing into individual GPU API calls
          * within the region if they require it (e.g. on OpenCL).

diff --git a/src/gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_types.h b/src/gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_types.h
index ceb3a3dafdf5c01b8f0df8b520844d5dd18f8cd8..0e0be3c1e1452fe9a2907880a70e7dee997a2201 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_types.h
+++ b/src/gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_types.h
@@ -276,7 +276,7 @@ struct gmx_nbnxn_cuda_t
      * setting bDoTime needs to be change if this CUDA "feature" gets fixed. */
     bool                 bDoTime;   /**< True if event-based timing is enabled.               */
     cu_timers_t         *timers;    /**< CUDA event-based timers.                             */
-    gmx_wallclock_gpu_t *timings;   /**< Timing data.                                         */
+    gmx_wallclock_gpu_t *timings;   /**< Timing data. TODO: deprecate this and query timers for accumulated data instead */
 };
 
 #ifdef __cplusplus

diff --git a/src/gromacs/mdlib/nbnxn_ocl/nbnxn_ocl_types.h b/src/gromacs/mdlib/nbnxn_ocl/nbnxn_ocl_types.h
index 74e98db9e322b237fe537f1281aab3963b1f7d18..0e9d527cee14987f7c1a1237e62d3f972de1393c 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_ocl/nbnxn_ocl_types.h
+++ b/src/gromacs/mdlib/nbnxn_ocl/nbnxn_ocl_types.h
@@ -368,7 +368,7 @@ struct gmx_nbnxn_ocl_t
 
     cl_bool                     bDoTime;  /**< True if event-based timing is enabled.                     */
     cl_timers_t                *timers;   /**< OpenCL event-based timers.                                 */
-    struct gmx_wallclock_gpu_t *timings;  /**< Timing data.                                               */
+    struct gmx_wallclock_gpu_t *timings;  /**< Timing data. TODO: deprecate this and query timers for accumulated data instead */
 };
 
 #ifdef __cplusplus