PME GPU/CUDA data framework.

[alexxy/gromacs.git] / src / gromacs / timing / gpu_timing.h

diff --git a/src/gromacs/timing/gpu_timing.h b/src/gromacs/timing/gpu_timing.h
index e9ec67d8555e864cdb46ea6918c4a7175b83b327..214b0120eb11505e685eedf054fedbfb990870ba 100644 (file)
--- a/src/gromacs/timing/gpu_timing.h
+++ b/src/gromacs/timing/gpu_timing.h
@@ -1,7 +1,7 @@
 /*
  * This file is part of the GROMACS molecular simulation package.
  *
- * Copyright (c) 2012,2014,2015,2017, by the GROMACS development team, led by
+ * Copyright (c) 2012,2014,2015,2016,2017, by the GROMACS development team, led by
  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
  * and including many others, as listed in the AUTHORS file in the
  * top-level source directory and at http://www.gromacs.org.
@@ -47,26 +47,52 @@
 extern "C" {
 #endif
 
-/*! \internal \brief Nonbonded kernel time and call count. */
-struct gmx_nbnxn_kernel_timing_data_t
+/*! \internal \brief GPU kernel time and call count. */
+struct gmx_kernel_timing_data_t
 {
     double  t; /**< Accumulated lapsed time */
     int     c; /**< Number of calls corresponding to the elapsed time */
 };
 
-/*! \internal \brief GPU timings for kernels and H2d/D2H transfers. */
-struct gmx_wallclock_gpu_t
+/*! \internal \brief
+ *
+ * PME GPU stages timing events indices, corresponding to the string in PMEStageNames in wallcycle.cpp.
+ */
+enum
+{
+    gtPME_SPLINE = 0,
+    gtPME_SPREAD,
+    gtPME_SPLINEANDSPREAD,
+    gtPME_FFT_R2C,
+    gtPME_SOLVE,
+    gtPME_FFT_C2R,
+    gtPME_GATHER,
+    gtPME_EVENT_COUNT /* not a stage ID but a static array size */
+};
+
+/*! \internal \brief GPU timings for PME. */
+struct gmx_wallclock_gpu_pme_t
+{
+    /* A separate PME structure to avoid refactoring the NB code for gmx_wallclock_gpu_t later
+     * TODO: devise a better GPU timing data structuring.
+     */
+    /*! \brief Array of PME GPU timing data. */
+    gmx_kernel_timing_data_t timing[gtPME_EVENT_COUNT];
+};
+
+/*! \internal \brief GPU NB timings for kernels and H2d/D2H transfers. */
+struct gmx_wallclock_gpu_nbnxn_t
 {
-    gmx_nbnxn_kernel_timing_data_t ktime[2][2];      /**< table containing the timings of the four
+    gmx_kernel_timing_data_t ktime[2][2];      /**< table containing the timings of the four
                                                         versions of the nonbonded kernels: force-only,
                                                         force+energy, force+pruning, and force+energy+pruning */
-    gmx_nbnxn_kernel_timing_data_t pruneTime;        /**< table containing the timings of the 1st pass prune-only kernels */
-    gmx_nbnxn_kernel_timing_data_t dynamicPruneTime; /**< table containing the timings of dynamic prune-only kernels */
-    double                         nb_h2d_t;         /**< host to device transfer time in nb calculation  */
-    double                         nb_d2h_t;         /**< device to host transfer time in nb calculation */
-    int                            nb_c;             /**< total call count of the nonbonded gpu operations */
-    double                         pl_h2d_t;         /**< pair search step host to device transfer time */
-    int                            pl_h2d_c;         /**< pair search step  host to device transfer call count */
+    gmx_kernel_timing_data_t pruneTime;        /**< table containing the timings of the 1st pass prune-only kernels */
+    gmx_kernel_timing_data_t dynamicPruneTime; /**< table containing the timings of dynamic prune-only kernels */
+    double                   nb_h2d_t;         /**< host to device transfer time in nb calculation  */
+    double                   nb_d2h_t;         /**< device to host transfer time in nb calculation */
+    int                      nb_c;             /**< total call count of the nonbonded gpu operations */
+    double                   pl_h2d_t;         /**< pair search step host to device transfer time */
+    int                      pl_h2d_c;         /**< pair search step  host to device transfer call count */
 };
 
 #ifdef __cplusplus