Pipeline GPU PME Spline/Spread with PP Comms

[alexxy/gromacs.git] / src / gromacs / ewald / pme_gpu_types.h

diff --git a/src/gromacs/ewald/pme_gpu_types.h b/src/gromacs/ewald/pme_gpu_types.h
index 97f84164996cd8c5ea685e26f790250194ef5129..76e111e7656dee2425874980843a6671d559fa89 100644 (file)
--- a/src/gromacs/ewald/pme_gpu_types.h
+++ b/src/gromacs/ewald/pme_gpu_types.h
@@ -1,7 +1,7 @@
 /*
  * This file is part of the GROMACS molecular simulation package.
  *
- * Copyright (c) 2016,2017,2018,2019,2020, by the GROMACS development team, led by
+ * Copyright (c) 2016,2017,2018,2019,2020,2021, 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.
@@ -82,6 +82,11 @@ static_assert(sizeof(DeviceBuffer<int>) == 8,
 #    define HIDE_FROM_OPENCL_COMPILER(x) char8
 #endif
 
+#ifndef NUMFEPSTATES
+//! Number of FEP states.
+#    define NUMFEPSTATES 2
+#endif
+
 /* What follows is all the PME GPU function arguments,
  * sorted into several device-side structures depending on the update rate.
  * This is GPU agnostic (float3 replaced by float[3], etc.).
@@ -95,11 +100,11 @@ static_assert(sizeof(DeviceBuffer<int>) == 8,
  */
 struct PmeGpuConstParams
 {
-    /*! \brief Electrostatics coefficient = ONE_4PI_EPS0 / pme->epsilon_r */
+    /*! \brief Electrostatics coefficient = c_one4PiEps0 / pme->epsilon_r */
     float elFactor;
     /*! \brief Virial and energy GPU array. Size is c_virialAndEnergyCount (7) floats.
      * The element order is virxx, viryy, virzz, virxy, virxz, viryz, energy. */
-    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_virialAndEnergy;
+    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_virialAndEnergy[NUMFEPSTATES];
 };
 
 /*! \internal \brief
@@ -130,14 +135,14 @@ struct PmeGpuGridParams
 
     /* Grid arrays */
     /*! \brief Real space grid. */
-    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_realGrid;
+    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_realGrid[NUMFEPSTATES];
     /*! \brief Complex grid - used in FFT/solve. If inplace cu/clFFT is used, then it is the same handle as realGrid. */
-    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_fourierGrid;
+    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_fourierGrid[NUMFEPSTATES];
 
     /*! \brief Grid spline values as in pme->bsp_mod
      * (laid out sequentially (XXX....XYYY......YZZZ.....Z))
      */
-    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_splineModuli;
+    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_splineModuli[NUMFEPSTATES];
     /*! \brief Fractional shifts lookup table as in pme->fshx/fshy/fshz, laid out sequentially (XXX....XYYY......YZZZ.....Z) */
     HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_fractShiftsTable;
     /*! \brief Gridline indices lookup table
@@ -158,15 +163,15 @@ struct PmeGpuAtomParams
      * but reallocation happens only at DD.
      */
     HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<gmx::RVec>) d_coordinates;
-    /*! \brief Global GPU memory array handle with input atom charges.
+    /*! \brief Global GPU memory array handle with input atom charges in states A and B.
      * The charges only need to be reallocated and copied to the GPU at DD step.
      */
-    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_coefficients;
+    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_coefficients[NUMFEPSTATES];
     /*! \brief Global GPU memory array handle with input/output rvec atom forces.
      * The forces change and need to be copied from (and possibly to) the GPU for every PME
      * computation, but reallocation happens only at DD.
      */
-    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<float>) d_forces;
+    HIDE_FROM_OPENCL_COMPILER(DeviceBuffer<gmx::RVec>) d_forces;
     /*! \brief Global GPU memory array handle with ivec atom gridline indices.
      * Computed on GPU in the spline calculation part.
      */
@@ -196,6 +201,9 @@ struct PmeGpuDynamicParams
     float recipBox[DIM][DIM];
     /*! \brief The unit cell volume for solving. */
     float boxVolume;
+
+    /*! \brief The current coefficient scaling value. */
+    float scale;
 };
 
 /*! \internal \brief
@@ -216,6 +224,15 @@ struct PmeGpuKernelParamsBase
      * before launching spreading.
      */
     struct PmeGpuDynamicParams current;
+
+    /*! \brief Whether pipelining with PP communications is active
+     * char rather than bool to avoid problem with OpenCL compiler */
+    char usePipeline;
+    /*! \brief Start atom for this stage of pipeline */
+    int pipelineAtomStart;
+    /*! \brief End atom for this stage of pipeline */
+    int pipelineAtomEnd;
+
     /* These texture objects are only used in CUDA and are related to the grid size. */
     /*! \brief Texture object for accessing grid.d_fractShiftsTable */
     HIDE_FROM_OPENCL_COMPILER(DeviceTexture) fractShiftsTableTexture;