* Copyright (c) 1991-2000, University of Groningen, The Netherlands.
* Copyright (c) 2001-2004, The GROMACS development team.
* Copyright (c) 2013,2014,2015,2016,2017 by the GROMACS development team.
- * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
+ * Copyright (c) 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.
#define GMX_EWALD_PME_H
#include <string>
+#include <vector>
#include "gromacs/gpu_utils/devicebuffer_datatype.h"
#include "gromacs/gpu_utils/gpu_macros.h"
#include "gromacs/math/vectypes.h"
-#include "gromacs/utility/arrayref.h"
-#include "gromacs/utility/basedefinitions.h"
#include "gromacs/utility/real.h"
struct gmx_hw_info_t;
struct t_nrnb;
struct PmeGpu;
struct gmx_wallclock_gpu_pme_t;
-struct DeviceInformation;
struct gmx_enerdata_t;
struct gmx_mtop_t;
struct gmx_pme_t;
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>
+class ArrayRef;
class ForceWithVirial;
class MDLogger;
enum class PinningPolicy : int;
class StepWorkload;
+
+/*! \libinternal \brief Class for managing usage of separate PME-only ranks
+ *
+ * Used for checking if some parts of the code could not use PME-only ranks
+ *
+ */
+class SeparatePmeRanksPermitted
+{
+public:
+ //! Disables PME ranks permitted flag with a reason
+ void disablePmeRanks(const std::string& reason);
+ //! Return status of PME ranks usage
+ bool permitSeparatePmeRanks() const;
+ //! Returns all reasons, for not using PME ranks
+ std::string reasonsWhyDisabled() const;
+
+private:
+ //! Flag that informs whether simualtion could use dedicated PME ranks
+ bool permitSeparatePmeRanks_ = true;
+ //! Storage for all reasons, why PME ranks could not be used
+ std::vector<std::string> reasons_;
+};
+
+class PmeCoordinateReceiverGpu;
} // namespace gmx
enum
* \todo We should evolve something like a \c GpuManager that holds \c
* DeviceInformation* and \c PmeGpuProgram* and perhaps other
* related things whose lifetime can/should exceed that of a task (or
- * perhaps task manager). See Redmine #2522.
+ * perhaps task manager). See Issue #2522.
*/
-gmx_pme_t* gmx_pme_init(const t_commrec* cr,
- const NumPmeDomains& numPmeDomains,
- const t_inputrec* ir,
- gmx_bool bFreeEnergy_q,
- gmx_bool bFreeEnergy_lj,
- gmx_bool bReproducible,
- real ewaldcoeff_q,
- real ewaldcoeff_lj,
- int nthread,
- PmeRunMode runMode,
- PmeGpu* pmeGpu,
- const DeviceInformation* deviceInfo,
- const PmeGpuProgram* pmeGpuProgram,
- const gmx::MDLogger& mdlog);
+gmx_pme_t* gmx_pme_init(const t_commrec* cr,
+ const NumPmeDomains& numPmeDomains,
+ const t_inputrec* ir,
+ gmx_bool bFreeEnergy_q,
+ gmx_bool bFreeEnergy_lj,
+ gmx_bool bReproducible,
+ real ewaldcoeff_q,
+ real ewaldcoeff_lj,
+ int nthread,
+ PmeRunMode runMode,
+ PmeGpu* pmeGpu,
+ const DeviceContext* deviceContext,
+ const DeviceStream* deviceStream,
+ const PmeGpuProgram* pmeGpuProgram,
+ const gmx::MDLogger& mdlog);
/*! \brief As gmx_pme_init, but takes most settings, except the grid/Ewald coefficients, from
* pme_src. This is only called when the PME cut-off/grid size changes.
int gmx_pme_do(struct gmx_pme_t* pme,
gmx::ArrayRef<const gmx::RVec> coordinates,
gmx::ArrayRef<gmx::RVec> forces,
- real chargeA[],
- real chargeB[],
- real c6A[],
- real c6B[],
- real sigmaA[],
- real sigmaB[],
+ gmx::ArrayRef<const real> chargeA,
+ gmx::ArrayRef<const real> chargeB,
+ gmx::ArrayRef<const real> c6A,
+ gmx::ArrayRef<const real> c6B,
+ gmx::ArrayRef<const real> sigmaA,
+ gmx::ArrayRef<const real> sigmaB,
const matrix box,
const t_commrec* cr,
int maxshift_x,
* 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.).
*
* \param[in,out] pme The PME structure.
* \param[in] numAtoms The number of particles.
- * \param[in] charges The pointer to the array of particle charges.
+ * \param[in] chargesA The pointer to the array of particle charges in the normal state or FEP
+ * state A. Can be nullptr if PME is not performed on the GPU.
+ * \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* charges);
+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 */
* formed gmx_pme_t structure. Should that one go away/work with inputrec?
*
* \param[in] ir Input system.
- * \param[in] mtop Complete system topology to check if an FE simulation perturbs charges.
* \param[out] error If non-null, the error message if the input is not supported on GPU.
*
* \returns true if PME can run on GPU with this input, false otherwise.
*/
-bool pme_gpu_supports_input(const t_inputrec& ir, const gmx_mtop_t& mtop, 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).
/*! \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] 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)) 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.
/*! \brief
* Launches last stage of PME on GPU - force gathering and D2H force transfer.
*
- * \param[in] pme The PME data structure.
- * \param[in] wcycle The wallclock counter.
+ * \param[in] pme The PME data structure.
+ * \param[in] wcycle The wallclock counter.
+ * \param[in] lambdaQ The Coulomb lambda to use when calculating the results.
*/
GPU_FUNC_QUALIFIER void pme_gpu_launch_gather(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
- gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle)) GPU_FUNC_TERM;
+ gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
+ real GPU_FUNC_ARGUMENT(lambdaQ)) GPU_FUNC_TERM;
/*! \brief
* Attempts to complete PME GPU tasks.
* \param[in] wcycle The wallclock counter.
* \param[out] forceWithVirial The output force and virial
* \param[out] enerd The output energies
+ * \param[in] lambdaQ The Coulomb lambda to use when calculating the results.
* \param[in] completionKind Indicates whether PME task completion should only be checked rather
* than waited for
* \returns True if the PME GPU tasks have completed
gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd),
+ real GPU_FUNC_ARGUMENT(lambdaQ),
GpuTaskCompletion GPU_FUNC_ARGUMENT(completionKind))
GPU_FUNC_TERM_WITH_RETURN(false);
* \param[in] wcycle The wallclock counter.
* \param[out] forceWithVirial The output force and virial
* \param[out] enerd The output energies
+ * \param[in] lambdaQ The Coulomb lambda to use when calculating the results.
*/
GPU_FUNC_QUALIFIER void pme_gpu_wait_and_reduce(gmx_pme_t* GPU_FUNC_ARGUMENT(pme),
const gmx::StepWorkload& GPU_FUNC_ARGUMENT(stepWork),
gmx_wallcycle* GPU_FUNC_ARGUMENT(wcycle),
gmx::ForceWithVirial* GPU_FUNC_ARGUMENT(forceWithVirial),
- gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd)) GPU_FUNC_TERM;
+ gmx_enerdata_t* GPU_FUNC_ARGUMENT(enerd),
+ real GPU_FUNC_ARGUMENT(lambdaQ)) GPU_FUNC_TERM;
/*! \brief
* The PME GPU reinitialization function that is called both at the end of any PME computation and on any load balancing.
* \param[in] pme The PME data structure.
* \returns Pointer to force data
*/
-GPU_FUNC_QUALIFIER void* pme_gpu_get_device_f(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
- GPU_FUNC_TERM_WITH_RETURN(nullptr);
-
-/*! \brief Returns the pointer to the GPU stream.
- * \param[in] pme The PME data structure.
- * \returns Pointer to GPU stream object.
- */
-GPU_FUNC_QUALIFIER const DeviceStream* pme_gpu_get_device_stream(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
- GPU_FUNC_TERM_WITH_RETURN(nullptr);
+GPU_FUNC_QUALIFIER DeviceBuffer<gmx::RVec> pme_gpu_get_device_f(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
+ GPU_FUNC_TERM_WITH_RETURN(DeviceBuffer<gmx::RVec>{});
/*! \brief Get pointer to the device synchronizer object that allows syncing on PME force calculation completion
* \param[in] pme The PME data structure.
- * \returns Pointer to sychronizer
+ * \returns Pointer to synchronizer
*/
GPU_FUNC_QUALIFIER GpuEventSynchronizer* pme_gpu_get_f_ready_synchronizer(const gmx_pme_t* GPU_FUNC_ARGUMENT(pme))
GPU_FUNC_TERM_WITH_RETURN(nullptr);
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff