Merge release-2021 into master

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

diff --git a/src/gromacs/ewald/pme.h b/src/gromacs/ewald/pme.h
index d897c0fb0caf08004fa44908da06c5738d09d859..eb0c5651753535db601959d2101f7f8e12669ebd 100644 (file)
--- a/src/gromacs/ewald/pme.h
+++ b/src/gromacs/ewald/pme.h
@@ -55,7 +55,6 @@
 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
 #include "gromacs/gpu_utils/gpu_macros.h"
 #include "gromacs/math/vectypes.h"
-#include "gromacs/utility/basedefinitions.h"
 #include "gromacs/utility/real.h"
 
 struct gmx_hw_info_t;
@@ -76,6 +75,19 @@ enum class GpuTaskCompletion;
 class PmeGpuProgram;
 class GpuEventSynchronizer;
 
+/*! \brief Hack to selectively enable some parts of PME during unit testing.
+ *
+ * Set to \c false by default. If any of the tests sets it to \c true, it will
+ * make the compatibility check consider PME to be supported in SYCL builds.
+ *
+ * Currently we don't have proper PME implementation with SYCL, but we still want
+ * to run tests for some of the kernels.
+ *
+ * \todo Remove after #3927 is done and PME is fully enabled in SYCL builds.
+ */
+//NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
+extern bool g_allowPmeWithSyclForTesting;
+
 namespace gmx
 {
 template<typename>
@@ -107,6 +119,7 @@ private:
     std::vector<std::string> reasons_;
 };
 
+class PmeCoordinateReceiverGpu;
 } // namespace gmx
 
 enum
@@ -234,7 +247,7 @@ int gmx_pme_do(struct gmx_pme_t*              pme,
  * pme struct. Currently does not work in parallel or with free
  * energy.
  */
-void gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q, real* V);
+real gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q);
 
 /*! \brief
  * This function updates the local atom data on GPU after DD (charges, coordinates, etc.).
@@ -248,7 +261,10 @@ void gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::
  * \param[in]     chargesB   The pointer to the array of particle charges in state B. Only used if
  * charges are perturbed and can otherwise be nullptr.
  */
-void gmx_pme_reinit_atoms(gmx_pme_t* pme, int numAtoms, const real* chargesA, const real* chargesB);
+void gmx_pme_reinit_atoms(gmx_pme_t*                pme,
+                          int                       numAtoms,
+                          gmx::ArrayRef<const real> chargesA,
+                          gmx::ArrayRef<const real> chargesB);
 
 /* A block of PME GPU functions */
 
@@ -284,6 +300,17 @@ bool pme_gpu_supports_hardware(const gmx_hw_info_t& hwinfo, std::string* error);
  */
 bool pme_gpu_supports_input(const t_inputrec& ir, std::string* error);
 
+/*! \brief Checks whether the input system allows to run PME on GPU in Mixed mode.
+ * Assumes that the input system is compatible with GPU PME otherwise, that is,
+ * before calling this function one should check that \ref pme_gpu_supports_input returns \c true.
+ *
+ * \param[in]  ir     Input system.
+ * \param[out] error  If non-null, the error message if the input is not supported.
+ *
+ * \returns true if PME can run on GPU in Mixed mode with this input, false otherwise.
+ */
+bool pme_gpu_mixed_mode_supports_input(const t_inputrec& ir, std::string* error);
+
 /*! \brief
  * Returns the active PME codepath (CPU, GPU, mixed).
  * \todo This is a rather static data that should be managed by the higher level task scheduler.
@@ -358,17 +385,24 @@ GPU_FUNC_QUALIFIER void pme_gpu_prepare_computation(gmx_pme_t*     GPU_FUNC_ARGU
 /*! \brief
  * Launches first stage of PME on GPU - spreading kernel.
  *
- * \param[in] pme                The PME data structure.
- * \param[in] xReadyOnDevice     Event synchronizer indicating that the coordinates
- * are ready in the device memory; nullptr allowed only on separate PME ranks.
- * \param[in] wcycle             The wallclock counter.
- * \param[in] lambdaQ            The Coulomb lambda of the current state of the
- * system. Only used if FEP of Coulomb is active.
+ * \param[in] pme                            The PME data structure.
+ * \param[in] xReadyOnDevice                 Event synchronizer indicating that the coordinates
+ *                                           are ready in the device memory; nullptr allowed only
+ *                                           on separate PME ranks.
+ * \param[in] wcycle                         The wallclock counter.
+ * \param[in] lambdaQ                        The Coulomb lambda of the current state of the
+ *                                           system. Only used if FEP of Coulomb is active.
+ * \param[in] useGpuDirectComm               Whether direct GPU PME-PP communication is active
+ * \param[in]  pmeCoordinateReceiverGpu      Coordinate receiver object, which must be valid when
+ *                                           direct GPU PME-PP communication is active
  */
-GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(gmx_pme_t*            GPU_FUNC_ARGUMENT(pme),
-                                              GpuEventSynchronizer* GPU_FUNC_ARGUMENT(xReadyOnDevice),
-                                              gmx_wallcycle*        GPU_FUNC_ARGUMENT(wcycle),
-                                              real GPU_FUNC_ARGUMENT(lambdaQ)) GPU_FUNC_TERM;
+GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(
+        gmx_pme_t*                     GPU_FUNC_ARGUMENT(pme),
+        GpuEventSynchronizer*          GPU_FUNC_ARGUMENT(xReadyOnDevice),
+        gmx_wallcycle*                 GPU_FUNC_ARGUMENT(wcycle),
+        real                           GPU_FUNC_ARGUMENT(lambdaQ),
+        bool                           GPU_FUNC_ARGUMENT(useGpuDirectComm),
+        gmx::PmeCoordinateReceiverGpu* GPU_FUNC_ARGUMENT(pmeCoordinateReceiverGpu)) GPU_FUNC_TERM;
 
 /*! \brief
  * Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.